State of AI Report 2022 - ONLINE.pdf

State of AI Report
October 11, 2022
#stateofai
stateof.ai
Ian Hogarth
Nathan Benaich

About the authors
Nathan is the General Partner of Air Street Capital, a
venture capital ﬁrm investing in AI-ﬁrst technology
and life science companies. He founded RAAIS and
London.AI (AI community for industry and research),
the RAAIS Foundation (funding open-source AI
projects), and Spinout.fyi (improving university spinout
creation). He studied biology at Williams College and
earned a PhD from Cambridge in cancer research.
Nathan Benaich Ian Hogarth
Ian is a co-founder at Plural, an investment platform
for experienced founders to help the most ambitious
European startups. He is a Visiting Professor at UCL
working with Professor Mariana Mazzucato. Ian was
co-founder and CEO of Songkick, the concert service.
He started studying machine learning in 2005 where
his Masters project was a computer vision system to
classify breast cancer biopsy images.
Introduction | Research | Industry | Politics | Safety | Predictions #stateofai | 2
stateof.ai 2022

Artificial intelligence (AI) is a multidisciplinary field of science and engineering whose goal is to create intelligent machines.
We believe that AI will be a force multiplier on technological progress in our increasingly digital, data-driven world. This is
because everything around us today, ranging from culture to consumer products, is a product of intelligence.
The State of AI Report is now in its fifth year. Consider this report as a compilation of the most interesting things we’ve seen
with a goal of triggering an informed conversation about the state of AI and its implication for the future.
We consider the following key dimensions in our report:
- Research: Technology breakthroughs and their capabilities.
- Industry: Areas of commercial application for AI and its business impact.
- Politics: Regulation of AI, its economic implications and the evolving geopolitics of AI.
- Safety: Identifying and mitigating catastrophic risks that highly-capable future AI systems could pose to us.
- Predictions: What we believe will happen in the next 12 months and a 2021 performance review to keep us honest.
Produced by Nathan Benaich (@nathanbenaich), Ian Hogarth (@soundboy), Othmane Sebbouh (@osebbouh) and
Nitarshan Rajkumar (@nitarshan).
stateof.ai 2022
#stateofai | 3
Introduction | Research | Industry | Politics | Safety | Predictions

Thank you!
Othmane Sebbouh
Research Assistant
Othmane is a PhD student in ML at
ENS Paris, CREST-ENSAE and CNRS.
He holds an MsC in management
from ESSEC Business School and a
Master in Applied Mathematics from
ENSAE and Ecole Polytechnique.
#stateofai | 4
Nitarshan Rajkumar
Research Assistant
Nitarshan is a PhD student in AI at
the University of Cambridge. He was
a research student at Mila and a
software engineer at Airbnb. He
holds a BSc from University of
Waterloo.
stateof.ai 2022

Definitions
stateof.ai 2022
#stateofai | 5
Artificial intelligence (AI): a broad discipline with the goal of creating intelligent machines, as opposed to the natural
intelligence that is demonstrated by humans and animals.
Artificial general intelligence (AGI): a term used to describe future machines that could match and then exceed the full
range of human cognitive ability across all economically valuable tasks.
AI Safety: a field that studies and attempts to mitigate the catastrophic risks which future AI could pose to humanity.
Machine learning (ML): a subset of AI that often uses statistical techniques to give machines the ability to "learn" from data
without being explicitly given the instructions for how to do so. This process is known as “training” a “model” using a learning
“algorithm” that progressively improves model performance on a specific task.
Reinforcement learning (RL): an area of ML in which software agents learn goal-oriented behavior by trial and error in an
environment that provides rewards or penalties in response to their actions (called a “policy”) towards achieving that goal.
Deep learning (DL): an area of ML that attempts to mimic the activity in layers of neurons in the brain to learn how to
recognise complex patterns in data. The “deep” refers to the large number of layers of neurons in contemporary models that
help to learn rich representations of data to achieve better performance gains.

Definitions
stateof.ai 2022
#stateofai | 6
Model: once a ML algorithm has been trained on data, the output of the process is known as the model. This can then be
used to make predictions.
Self-supervised learning (SSL): a form of unsupervised learning, where manually labeled data is not needed. Raw data is
instead modified in an automated way to create artificial labels to learn from. An example of SSL is learning to complete text
by masking random words in a sentence and trying to predict the missing ones.
(Large) Language model (LM, LLM): a model trained on textual data. The most common use case of a LM is text generation.
The term “LLM” is used to designate multi-billion parameter LMs, but this is a moving definition.
Computer vision (CV): enabling machines to analyse, understand and manipulate images and video.
Transformer: a model architecture at the core of most state of the art (SOTA) ML research. It is composed of multiple
“attention” layers which learn which parts of the input data are the most important for a given task. Transformers started in
language modeling, then expanded into computer vision, audio, and other modalities.

Research
- Diffusion models took the computer vision world by storm with impressive text-to-image generation capabilities.
- AI attacks more science problems, ranging from plastic recycling, nuclear fusion reactor control, and natural product discovery.
- Scaling laws refocus on data: perhaps model scale is not all that you need. Progress towards a single model to rule them all.
- Community-driven open sourcing of large models happens at breakneck speed, empowering collectives to compete with large labs.
- Inspired by neuroscience, AI research are starting to look like cognitive science in its approaches.
Industry
- Have upstart AI semiconductor startups made a dent vs. NVIDIA? Usage statistics in AI research shows NVIDIA ahead by 20-100x.
- Big tech companies expand their AI clouds and form large partnerships with A(G)I startups.
- Hiring freezes and the disbanding of AI labs precipitates the formation of many startups from giants including DeepMind and OpenAI.
- Major AI drug discovery companies have 18 clinical assets and the ﬁrst CE mark is awarded for autonomous medical imaging diagnostics.
- The latest in AI for code research is quickly translated by big tech and startups into commercial developer tools.
Politics
- The chasm between academia and industry in large scale AI work is potentially beyond repair: almost 0% of work is done in academia.
- Academia is passing the baton to decentralized research collectives funded by non-traditional sources.
- The Great Reshoring of American semiconductor capabilities is kicked off in earnest, but geopolitical tensions are sky high.
- AI continues to be infused into a greater number of defense product categories and defense AI startups receive even more funding.
Safety
- AI Safety research is seeing increased awareness, talent, and funding, but is still far behind that of capabilities research.
Executive Summary
stateof.ai 2022
#stateofai | 7

Scorecard: Reviewing our predictions from 2021
stateof.ai 2022
#stateofai | 8

Our 2021 Prediction Grade Evidence
Transformers replace RNNs to learn world models with which RL
agents surpass human performance in large and rich games.
Yes
DeepMind’s Gato model makes progress in this direction in which a transformer
predicts the next state and action, but it is not trained with RL. University of
Geneva’s GPT-like transformer model IRIS solves tasks in Atari environments.
ASML’s market cap reaches $500B. No Current market cap is circa $165B (3 Oct 2022)
Anthropic publishes on the level of GPT, Dota, AlphaGo to establish
itself as a third pole of AGI research.
No Not yet.
A wave of consolidation in AI semiconductors with at least one of
Graphcore, Cerebras, SambaNova, Groq, or Mythic being acquired by
a large technology company or major semiconductor incumbent.
No No new announced AI semiconductor consolidation has happened yet.
Small transformers + CNN hybrid models match current SOTA on
ImageNet top-1 accuracy (CoAtNet-7, 90.88%, 2.44B params) with
10x fewer parameters.
Yes
MaxViT from Google with 475M parameters almost matched (89.53%) CoAtNet-7’s
performance (90.88%) on ImageNet top-1 accuracy.
DeepMind shows a major breakthrough in the physical sciences. Yes Three (!) DeepMind papers in mathematics and material science.
The JAX framework grows from 1% to 5% of monthly repos created
as measured by Papers With Code.
No JAX usage still accounts for <1% of monthly repos on Papers With Code.
A new AGI-focused research company is formed with signiﬁcant
backing and a roadmap that’s focused on a sector vertical (e.g.
developer tools, life science).
Yes
Adept.ai was co-founded by the authors of the Transformer and is focused on AGI
via software tool use automation.
#stateofai | 9

#stateofai | 10
Bonus! Predictions, revisited - better late than never!
Year Prediction Grade Evidence
2018
Access to Taiwanese and South Korean semiconductor companies
becomes an explicit part of the trade war between US and China.
Yes
US CHIPS Act 2022 prevents recipients to expand
operations in China. TSMC caught in the crosshairs.
2018
The government of an OECD country blocks the acquisition of a
leading ML company by a US or Chinese HQ'd tech company.
Yes
The UK, amongst others, blocked the acquisition of Arm
by NVIDIA.
2019
As AI systems become more powerful, governance of AI becomes a
bigger topic and at least one major AI company makes a substantial
change to their governance model.
Yes Anthropic set up as a public beneﬁt corporation.
2020
Facebook/Meta makes a major breakthrough in AR/VR with 3D
computer vision.
Sort of Implicitron in PyTorch3D. Not applied to AR/VR yet.
2020
Chinese and European defense-focused AI startups collectively raise
>$100M in the next 12 months.
Yes Helsing (Germany) raised $100M Series A in 2022.
2020 NVIDIA does not end up completing its acquisition of Arm. Yes Deal is formally cancelled in 2022.
stateof.ai 2022

Section 1: Research
stateof.ai 2022
#stateofai | 11

In 2021, we predicted: “DeepMind releases a major research breakthrough in the physical sciences.” The company
has since made significant advancements in both mathematics and materials science.
stateof.ai 2022
● One of the decisive moments in mathematics is formulating
a conjecture, or a hypothesis, on the relationship between
variables of interest. This is often done by observing a large
number of instances of the values of these variables, and
potentially using data-driven conjecture generation methods.
But these are limited to low-dimensional, linear, and
generally simple mathematical objects.
#stateofai | 12
2021 Prediction: DeepMind’s breakthroughs in the physical sciences (1/3)
● In a Nature article, DeepMind researchers proposed an iterative workflow involving mathematicians and a
supervised ML model (typically a NN). Mathematicians hypothesize a function relating two variables (input X(z)
and output Y(z)). A computer generates a large number of instances of the variables and a NN is fit to the data.
Gradient saliency methods are used to determine the most relevant inputs in X(z). Mathematicians can turn
refine their hypothesis and/or generate more data until the conjecture holds on a large amount of data.

stateof.ai 2022
● DeepMind researchers used their framework in a collaboration with
mathematics professors from the University of Sydney and the University of
Oxford to (i) propose an algorithm that could solve a 40 years-long standing
conjecture in representation theory and (ii) prove a new theorem in the
study of knots.
#stateofai | 13
● DeepMind made an important contribution in materials science as
well. It showed that the exact functional in Density Functional
Theory, an essential tool to compute electronic energies, can be
efﬁciently approximated using a neural network. Notably, instead of
constraining the neural network to verify mathematical constraints
of the DFT functional, researchers simply incorporate them into the
training data to which they ﬁt the NN.

stateof.ai 2022
● DeepMind repurposed AlphaZero (their RL model trained to beat the best
human players of Chess, Go and Shogi) to do matrix multiplication. This
AlphaTensor model was able to ﬁnd new deterministic algorithms to
multiply two matrices. To use AlphaZero, the researchers recast the matrix
multiplication problem as a single-player game where each move
corresponds to an algorithm instruction and the goal is to zero-out a tensor
measuring how far from correct the predicted algorithm is.
● Finding faster matrix multiplication algorithms, a seemingly simple and
well-studied problem, has been stale for decades. DeepMind’s approach not
only helps speed up research in the ﬁeld, but also boosts matrix
multiplication based technology, that is AI, imaging, and essentially
everything happening on our phones.
#stateofai | 14

● A popular route to achieving nuclear fusion requires
confining extremely hot plasma for enough time
using a tokamak.
● A major obstacle is that the plasma is unstable, loses
heat and degrades materials when it touches the
tokamak’s walls. Stabilizing it requires tuning the
magnetic coils thousands of times per second.
● DeepMind’s deep RL system did just that: first in a
simulated environment and then when deployed in
the TCV in Lausanne. The system was also able to
shape the plasma in new ways, including making it
compatible with ITER’s design.
DeepMind trained a reinforcement learning system to adjust the magnetic coils of Lausanne’s TCV (Variable
Configuration tokamak).The system’s flexibility means it could also be used in ITER, the promising next
generation tokamak under construction in France.
stateof.ai 2022
#stateofai | 15
Reinforcement learning could be a core component of the next fusion breakthrough

stateof.ai 2022
Predicting the structure of the entire known proteome: what could this unlock next?
Since its open sourcing, DeepMind’s AlphaFold 2 has been used in hundreds of research papers.The company has
now deployed the system to predict the 3D structure of 200 million known proteins from plants, bacteria,animals
and other organisms.The extent of the downstream breakthroughs enabled by this technology - ranging from
drug discovery to basic science - will need a few years to materialize.
● There are 190k empirically determined 3D structures in the Protein Data
Bank today. These have been derived through X-Ray crystallography and
cryogenic electron microscopy.
● The ﬁrst release of AlphaFold DB in July 2021 included 1M predicted
protein structures.
● This new release 200x’s the database size. Over 500,000 researchers from
190 countries have made use of the database.
● AlphaFold mentions in AI research literature is growing massively and is
predicted to triple year on year (right chart).
#stateofai | 16

● This is because ESMFold doesn’t rely on the use of multiple
sequence alignments (MSA) and templates like AlphaFold 2 and
RoseTTAFold, and instead only uses protein sequences.
stateof.ai 2022
Researchers independently applied language models to the problems of protein generation and structure
prediction while scaling model parameter.They both report large beneﬁts from scaling their models.
Language models for proteins: a familiar story of open source and scaled models
● Salesforce researchers ﬁnd that scaling their LMs allows them to better capture the
training distribution of protein sequences (as measured by perplexity).
● Using the 6B param ProGen2, they generated proteins with similar folds to natural
proteins, but with a substantially different sequence identity. But to unlock the full
potential of scale, the authors insist that more emphasis be placed on data distribution.
#stateofai | 17
● Meta et al. introduced the ESM family of protein LMs, whose sizes range from 8M to 15B
(dubbed ESM-2) parameters. Using ESM-2, they build ESMFold to predict protein structure.
They show that ESMFold produces similar predictions to AlphaFold 2 and RoseTTAFold,
but is an order of magnitude faster.

stateof.ai 2022
Researchers used CRISPR-based endogenous tagging — modifying genes by illuminating specific aspects of the
proteins’ function — to determine protein localization in cells.They then used clustering algorithms to identify
protein communities and formulate mechanistic hypotheses on uncharacterized proteins.
OpenCell: understanding protein localization with a little help from machine learning
● An important goal of genomic research is to understand where proteins localize and
how they interact in a cell to enable particular functions. With its dataset of 1,310
tagged proteins across ~5,900 3D images, the OpenCell initiative enabled
researchers to draw important links between spatial distribution of proteins, their
functions, and their interactions.
● Markov clustering on the graph of protein interactions successfully delineated
functionally related proteins. This will help researchers better understand so-far
uncharacterized proteins.
● We often expect ML to deliver definitive predictions. But here as with math, ML first
gives partial answers (here clusters), humans then interpret, formulate and test
hypotheses, before delivering a definitive answer.
#stateofai | 18

stateof.ai 2022
Researchers from UT Austin engineered an enzyme capable of degrading PET, a type of plastic responsible for
12% of global solid waste.
Plastic recycling gets a much-needed ML-engineered enzyme
● The PET hydrolase, called FAST-PETase, is more
robust to different temperatures and pH levels
than existing ones.
● FAST-PETase was able to almost completely
degrade 51 different products in 1 week.
#stateofai | 19
● They also showed that they
could resynthesize PET from
monomers recovered from
FAST-PETase degradation,
potentially opening the way
for industrial scale
closed-loop PET recycling.

stateof.ai 2022
With the increased use of ML in quantitative sciences, methodological errors in ML can leak to these disciplines.
Researchers from Princeton warn of a growing reproducibility crisis in ML-based science driven in part by one
such methodological error: data leakage.
Beware of compounded errors: in science, ML in and garbage out?
● Data leakage is an umbrella term covering all cases where data
that shouldn’t be available to a model in fact is. The most
common example is when test data is included in the training
set. But the leakage can be more pernicious: when the model
uses features that are a proxy of the outcome variable or when
test data come from a distribution which is different from the
one about which the scientific claim is made.
#stateofai | 20
● The authors argue that the ensuing reproducibility failures in ML-based science are systemic: they study 20
reviews across 17 science fields examining errors in ML-based science and find that data leakage errors
happened in every one of the 329 papers the reviews span. Inspired by the increasingly popular model cards in
ML, the authors propose that researchers use model info sheets designed to prevent data leakage issues.

● OpenAI gathered 2,000 hours of video labeled with
mouse and keyboard actions and trained an inverse
dynamics model (IDM) to predict actions given past
and future frames – this is the PreTraining part.
● They then used the IDM to label 70K hours of video
on which they trained a model to predict actions
given only past video frames.
● They show that the model can be fine-tuned with
imitation learning and reinforcement learning (RL)
to achieve a performance which is too hard to reach
using RL from scratch.
stateof.ai 2022
OpenAI trained a model (Video PreTraining, VPT) to play Minecraft from video frames using a small amount of
labeled mouse and keyboard interactions. VPT is the first ML model to learn to craft diamonds, “a task that usually
takes proficient humans over 20 minutes (24,000 actions)”.
#stateofai | 21
OpenAI uses Minecraft as a testbed for computer-using agents

stateof.ai 2022
OpenAI’s Codex, which drives GitHub Copilot, has impressed the computer science community with its ability to
complete code on multiple lines or directly from natural language instructions.This success spurred more
research in this space, including from Salesforce, Google and DeepMind.
● With the conversational CodeGen, Salesforce researchers leverage the language
understanding of LLMs to specify coding requirements in multiturn language
interactions. It is the only open source model to be competitive with Codex.
● A more impressive feat was achieved by Google’s LLM PaLM, which achieves a similar
performance to Codex, but with 50x less code in its training data (PaLM was trained
on a larger non-code dataset). When fine-tuned on Python code, PaLM outperformed
(82% vs. 71.7% SOTA) peers on Deepfix, a code repair task.
● DeepMind’s AlphaCode tackles a different problem: the generation of whole
programs on competitive programming tasks. It ranked in the top half on Codeforces,
a coding competitions platform. It was pre-trained on GitHub data and fine-tuned on
Codeforces problems and solutions. Millions of possible solutions are then sampled,
filtered, and clustered to obtain 10 final candidate submissions.
#stateofai | 22
Corporate AI labs rush into AI for code research

stateof.ai 2022
The attention layer at the core of the transformer model famously suffers from a quadratic dependence on its
input.A slew of papers promised to solve this, but no method has been adopted. SOTA LLMs come in different
flavors (autoencoding, autoregressive, encoder-decoders), yet all rely on the same attention mechanism.
Five years after the Transformer, there must be some efficient alternative, right… right?
● A Googol of transformers have been trained over the past few years,
costing millions (billions?) to labs and companies around the world.
But so-called “Efficient Transformers” are nowhere to be found in
large-scale LM research (where they would make the biggest
difference!). GPT-3, PaLM, LaMDA, Gopher, OPT, BLOOM, GPT-Neo,
Megatron-Turing NLG, GLM-130B, etc. all use the original attention
layer in their transformers.
● Several reasons can explain this lack of adoption: (i) the potential
linear speed-up is only useful for large input sequences, (ii) the new
methods introduce additional constraints that make the architectures
less universal, (iii) the reported efficiency measures don’t translate in
actual computational cost and time savings.
#stateofai | 23

stateof.ai 2022
Built on Google’s 540B parameter LM PaLM, Google’s Minerva achieves a 50.3% score on the MATH benchmark
(43.4 pct points better than previous SOTA), beating forecasters expectations for best score in 2022 (13%).
Meanwhile, OpenAI trained a network to solve two mathematical olympiad problems (IMO).
● Google trained its (pre-trained) LLM PaLM on an additional 118GB dataset of scientific
papers from arXiv and web pages using LaTeX and MathJax. Using chain of thought
prompting (including intermediate reasoning steps in prompts rather than the final
answer only) and other techniques like majority voting, Minerva improves the SOTA on
most datasets by at least double digit pct points.
● Minerva only uses a language model and doesn’t explicitly encode formal
mathematics. It is more flexible but can only be automatically evaluated on its final
answer rather than its whole reasoning, which might justify some score inflation.
● In contrast, OpenAI built a (transformer-based) theorem prover built in the Lean
formal environment. Different versions of their model were able to solve a number of
problems from AMC12 (26), AIME (6) and IMO (2) (increasing order of difficulty).
#stateofai | 24
Mathematical abilities of Language Models largely surpass expectations

stateof.ai 2022
Only 66% of machine learning benchmarks have received more than 3 results at different time points, and many
are solved or saturated soon after their release. BIG (Beyond the Imitation Game), a new benchmark designed by
444 authors across 132 institutions, aims to challenge current and future language models.
● A study from the University of Vienna, Oxford, and FHI
examined 1,688 benchmarks for 406 AI tasks and
identiﬁed different submission dynamics (see right).
● They note that language benchmarks in particular tend
to be quickly saturated.
#stateofai | 25
Fast progress in LLM research renders benchmarks obsolete, but a BIG one comes to help
● Rapid LLM progress and emerging capabilities seem to outrun current benchmarks. As a result, much of this
progress is only captured through circumstantial evidence like demos or one-off breakthroughs, and/or
evaluated on disparate dedicated benchmarks, making it difﬁcult to identify actual progress.
● The new BIG benchmark contains 204 tasks, all with strong human expert baselines, which evaluate a large set
of LLM capabilities from memorization to multi-step reasoning. They show that, for now, even the best models
perform poorly on the BIG benchmark.

stateof.ai 2022
DeepMind revisited LM scaling laws and found that current LMs are significantly undertrained: they’re not trained
on enough data given their large size. They train Chinchilla, a 4x smaller version of their Gopher, on 4.6x more
data, and find that Chinchilla outperforms Gopher and other large models on BIG-bench.
● Empirical LM scaling laws determine, for a fixed compute budget,
the model and training data sizes that should be used. Past work
from OpenAI had established that model size should increase faster
than training data size as the compute budget increases.
● DeepMind claims that the model size and the number of training
tokens should instead increase at roughly the same rate.
● Compared to OpenAI’s work, DeepMind uses larger models to derive
their scaling laws. They emphasize that data scaling leads to better
predictions from multibillion parameter models.
#stateofai | 26
Ducking language model scaling laws: more data please
● Following these new scaling laws, Chinchilla (70B params) is trained on 1.4T tokens. Gopher (230B) on 300B.
● Though trained with the same compute budget, the lighter Chinchilla should be faster to run.

● Emergence is not fully understood: it could be that for multi-step reasoning tasks,
models need to be deeper to encode the reasoning steps. For memorization tasks,
having more parameters is a natural solution. The metrics themselves may be part
of the explanation, as an answer on a reasoning task is only considered correct if its
conclusion is. Thus despite continuous improvements with model size, we only
consider a model successful when increments accumulate past a certain point.
● A possible consequence of emergence is that there are a range of tasks that are out
of reach of current LLMs that could soon be successfully tackled.
● Alternatively, deploying LLMs on real-world tasks at larger scales is more uncertain
as unsafe and undesirable abilities can emerge. Alongside the brittle nature of ML
models, this is another feature practitioners will need to account for.
While model loss can be reasonably predicted as a function of size and compute using well-calibrated scaling
laws, many LLM capabilities emerge unpredictably when models reach a critical size.These acquired capabilities
are exciting, but the emergence phenomenon makes evaluating model safety more difﬁcult.
stateof.ai 2022
#stateofai | 27
Ducking language model scaling laws: emergence
Arithmetics
Transliteration
Multi-task NLU
Fig. of speech
Training FLOPs

stateof.ai 2022
Language models can learn to use tools such as search engines and calculators, simply by making available text
interfaces to these tools and training on a very small number of human demonstrations.
● OpenAI’s WebGPT was the first model to demonstrate this convincingly by
fine-tuning GPT-3 to interact with a search engine to provide answers
grounded with references. This merely required collecting data of humans
doing this task and converting the interaction data into text that the
model could consume for training by standard supervised learning.
Importantly, the use of increasing amounts of human demonstration data
significantly increased the truthfulness and informativeness of answers
(right panel, white bars for WebGPT), a significant advance from when we
covered truthfulness evaluation in our 2021 report (slide 44).
● Adept, a new AGI company, is commercializing this paradigm. The company
trains large transformer models to interact with websites, software
applications and APIs (see more at adept.ai/act) in order to drive workflow
productivity.
#stateofai | 28
Teach a machine to fish: tool use as the next frontier?

stateof.ai 2022
A study documents the incredible acceleration of compute requirements in machine learning. It identiﬁes 3 eras
of machine learning according to training compute per model doubling time.The Pre-Deep Learning Era
(pre-2010, training compute doubled every 20 months), the Deep Learning Era (2010-15, doubling every 6
months), and the Large-Scale Era (2016-present, a 100-1000x jump, then doubling every 10 months).
#stateofai | 29
Looking back: three eras of compute in machine learning

stateof.ai 2022
When we covered diffusion models in the 2021 Report (slide 36), they were overtaking GANs in image generation
on a few benchmarks.Today, they are now the undisputable SOTA for text-to-image generation, and are diffusing
(pun intended) into text-to-video, text generation, audio, molecular design and more.
● Diffusion models (DMs) learn to reverse successive noise additions to images by
modeling the inverse distribution (generating denoised images from noisy ones) at each
step as a Gaussian whose mean and covariance are parametrized as a neural network.
DMs generate new images from random noise.
● Sequential denoising makes them slow, but new techniques (like denoising in a
lower-dimensional space) allow them to be faster at inference time and to generate
higher-quality samples (classifier-free guidance – trading off diversity for fidelity).
● SOTA text-to-image models like DALL-E 2, Imagen and Stable Diffusion are based on
DMs. They’re also used in controllable text generation (generating text with a pre-defined
structure or semantic context), model-based reinforcement learning, video generation
and even molecular generation.
#stateofai | 30
Diffusion models take over text-to-image generation and expand into other modalities

stateof.ai 2022
The second iteration of OpenAI’s DALL-E, released in April 2022, came with a signiﬁcant jump in the quality of
generated images. Soon after, another at least equally impressive diffusion-based model came from Google
(Imagen). Meanwhile, Google’s Parti took a different, autoregressive, route.
DALL-E 2, Imagen and Parti…the battle for text-to-image generation rages
● Instead of using a diffusion model, Parti treats text-to-image generation as a simple sequence-to-sequence
task, where the sequence to be predicted is a representation of the pixels of the image. Notably, as the number
of parameters and training data in Parti are scaled, the model acquires new abilities like spelling.
● Other impressive text-to-image models include GLIDE (OpenAI) and Make-a-Scene (Meta — can use both text
and sketches), which predate DALL-E 2, and CogView2 (Tsinghua, BAAI — both English and Chinese).
#stateofai | 31
DALL-E 2 Imagen Parti-350M Parti-20B

stateof.ai 2022
Stability.ai and Midjourney came out of seemingly nowhere with text-to-image models that rival those of
established AI labs. Both have APIs in beta, Midjourney is reportedly proﬁtable, and Stability has already
open-sourced their model. But more on their emergence and research dynamics in our Politics section.
#stateofai | 32
The text-to-image diffusion model frenzy gives birth to new AI labs
Image Credits: Fabian Stelzer

stateof.ai 2022
The text-to-video generation race has started
Research on diffusion-based text-to-video generation was kicked-off around April 2022, with work from Google
and the University of British Columbia. But in late September, new research from Meta and Google came with a
jump in quality, announcing a sooner-than-expected DALL-E moment for text-to-video generation.
● Meta made the ﬁrst splash from Big Tech in text-to-video generation by releasing Make-a-Video, a diffusion
model for video generation.
● In an eerily similar fashion to text-to-image generation, Google then published (less than a week later) almost
simultaneously two models: one diffusion-model based, Imagen, and another non diffusion-model based,
Phenaki. The latter can dynamically adapt the video via additional prompts.
#stateofai | 33
Imagen Video
Make-a-Video Phenaki

Closed for 14 months: community-driven open sourcing of GPT et al.
Landmark models from OpenAI and DeepMind have been implemented/cloned/improved by the open source
community much faster than we’d have expected.
stateof.ai 2022
#stateofai | 34
GPT-3 (175B) Pan-Gu (200B)
HyperCLOVA (204B)
Jurassic-1 Jumbo (204B)
FLAN (137B)
June 2020 May 2021
Megatron
Turing-NLG (137B)
Yuan 1.0 (246B)
Sep 2021
Gopher (280B)
Ernie 3.0 Titan (260B)
LaMDA (280B)
Jan 2022
GPT-j (6B)
GPT-NeoX (20B)
Aug 2021
PaLM (540B)
OPT (175B)
BLOOM (176B)
GLM (130B)
Open-sourced models in red
May 2022
Aug 2022
Chinchilla
(70B)

stateof.ai 2022
#stateofai | 35
DALL-E
Jan 2021
DALL-E 2
DALL-E mini
Apr 2022
Make-a-scene
Mar 2022 Aug 2022
May 2022 June 2022 July 2022
Imagen Parti Stable Diffusion
CogView2
Closed for 15 months: community-driven open sourcing of DALL-E et al.
Open-sourced models in red

stateof.ai 2022
#stateofai | 36
AlphaFold 1
Aug 2018
ESM
ESM 2
AlphaFold 2
July 2021
RosettaFold
Apr 2019
OpenFold
Aug 2022
Closed for 35 months: community-driven open sourcing of AlphaFold et al.
Open-sourced models in red.
Note that the models we reference are not necessarily replicas,but can be improved versions or independently developed.
ESM-1B
Nov 2020

stateof.ai 2022
Thanks to their large range of capabilities, LLMs could in principle enable robots to perform any task by
explaining its steps in natural language. But LLMs have little contextual knowledge of the robot’s environment
and its abilities, making their explanations generally infeasible for the robot. PaLM-SayCan solves this.
LLMs empower robots to execute diverse and ambiguous instructions
● Given an ambiguous instruction “I spilled my drink, can you help?”, a carefully
prompt-engineered LLM (e.g. Google’s PaLM) can devise a sequence of abstract steps to
pick up and bring you a sponge. But any given skill (e.g. pick up, put down) needs to be
doable by the robot in concordance with its environment (e.g. robot sees a sponge).
● To incentivise the LLM to output feasible instructions, SayCan maximises the likelihood
of an instruction being successfully executed by the robot.
● Assume the robot can execute a set of skills. Then, for any given instruction and state, the
system selects the skill that maximizes: the probability of a given completion (restricted
to the set of available skills) times the probability of success given the completion and
the current state. The system is trained using reinforcement learning.
● Researchers tested SayCan on 101 instructions from 7 types of language instructions. It
was successful in planning and execution 84% and 74% of the time respectively.
#stateofai | 37

stateof.ai 2022
The introduction of Vision Transformers (ViT) and other image transformers last year as SOTA models on imaging
benchmarks announced the dawn of ConvNets. Not so fast: work from Meta and UC Berkeley argues that
modernizing ConvNets gives them an edge over ViTs.
2021 Prediction: in vision, convolutional networks want a fair ﬁght with transformers…
● The researchers introduce ConvNeXt, a ResNet which is augmented
with the recent design choices introduced in hierarchical vision
Transformers like Swin, but doesn’t use attention layers.
● ConvNeXt is both competitive with Swin Transformer and ViT on
ImageNet-1K and ImageNet-22K and beneﬁts from scale like them.
● Transformers quickly replaced recurrent neural networks in language
modeling, but we don’t expect a similar abrupt drop-off in ConvNets
usage, especially in smaller scale ML use-cases.
● Meanwhile, our 2021 prediction of small transformers + CNN hybrid
models manifested in MaxViT from Google with 475M parameters
almost matching (89.53%) CoAtNet-7’s performance (90.88%) on
ImageNet top-1 accuracy.
#stateofai | 38

● Transformer-based autoencoder LMs are trained to predict randomly
masked words in large text corpora. This results in powerful models that
are SOTA in language modeling tasks (e.g. BERT).
● While masking a word in a sentence makes the sentence nonsensical
and creates a challenging task for LMs, reconstructing a few randomly
masked pixels in images is trivial thanks to neighbouring pixels.
● The solution: mask large patches of pixels (e.g. 75% of the pixels). Meta use this and other adjustments (the
encoder only sees visible patches, the decoder is much smaller than the encoder) to pre-train a ViT-Huge model
on ImageNet-1K and then fine-tune it to achieve a task-best 87.8% top-1 accuracy.
● Self-supervised learning isn’t new to computer vision (see for e.g. Meta’s SEER model). Nor are masking
techniques (e.g. Context encoders, or a more recent SiT). But this work is further evidence that SOTA techniques
in language transition seamlessly vision. Can domains unification be pushed further?
stateof.ai 2022
Self-supervision techniques used to train transformers on text are now transposed almost as is to images and are
achieving state of the art results on ImageNet-1K.
…but the inevitable vision and language modeling unification continues…
#stateofai | 39

stateof.ai 2022
Transformers trained on a specific task (via supervised or self-supervised learning) can be used for a broader set
of tasks via fine-tuning. Recent works show that a single transformer can be directly and efficiently trained on
various tasks across different modalities (multi-task multimodal learning).
● Attempts at generalist multitask, multimodal models date back to at
least Google’s “One model to learn them all” (2017), which tackled 8
tasks in image, text and speech. DeepMind’s Gato brings this effort to
another level: researchers train a 1.2B parameter transformer to
perform hundreds of tasks in robotics, simulated environments, and
vision and language. This partially proves our 2021 Prediction.
● They showed that scaling consistently improved the model, but it was
kept “small” for live low-latency robotics tasks.
#stateofai | 40
2021 Prediction:…culminating in a single transformer to rule them all?
● To train their model on different modalities, all data was serialized into a sequence of tokens which are
embedded in a learned vector space. The model is trained in a fully supervised fashion.
● Separately: With data2vec, on a narrower set of tasks, Meta devised a unified self-supervision strategy across
modalities. But for now, different transformers are used for each modality.

In 2021, we predicted: “Transformers replace RNNs to learn world models with which RL agents surpass human
performance in large and rich game environments.” Researchers from the University of Geneva used a GPT-like
transformer to simulate the world environment. They showed that their agent (dubbed IRIS) was sample efﬁcient
and surpassed human performance on 10 of the 26 games of Atari. IRIS was notably the best method among the
ones that don’t use lookahead search.
stateof.ai 2022
#stateofai | 41
2021 Prediction: transformers for learning in world models in reinforcement learning

In the 2020 State of AI Report we predicted that transformers would expand beyond NLP to achieve state of the
art in computer vision. It is now clear that transformers are a candidate general purpose architecture.Analysing
transformer-related papers in 2022 shows just how ubiquitous this model architecture has become.
stateof.ai 2022
#stateofai | 42
Transformers are becoming truly cross-modality
41%
22%
16%
9%
7%
5%
81%
2%

stateof.ai 2022
The seminal NeRF paper was published in March 2020. Since then, fundamental improvements to the methods
and new applications have been quickly and continuously developed. For example, more than 50 papers on NeRF
alone appeared at CVPR in 2022.
NeRFs expand into their own mature ﬁeld of research
● From last year’s Report (slide 18): Given multiple views of an image, NeRF uses a
multilayered perceptron to learn a representation of the image and to render
new views of it. It learns a mapping from every pixel location and view direction
to the color and density at that location.
● Among this year’s work, Plenoxels stands out by removing the MLP altogether
and achieving a 100x speedup in NeRF training. Another exciting direction was
rendering large scale sceneries from a few views with NeRFs, whether city-scale
(rendering entire neighborhoods of San Francisco with Block-NeRF) or
satellite-scale with Mega-NeRF*.
#stateofai | 43
● Given the current quality of the results and the ﬁeld’s rate of progress, we expect that in a year or two, NeRFs
will feature prominently in our industry section.
*You can better appreciate NeRF research by checking demos. E.g.
Block-NeRF, NeRF in the dark, Light Field Neural Rendering

stateof.ai 2022
Resistance to antibacterial agents is common and often arises as a result of a different pathogen already present
in a patient’s body. So how should doctors find the right antibiotic that cures the infection but doesn’t render the
patient susceptible to a new infection?
Treating bacterial infections by data-driven personalised selection of antibacterial agents
● By comparing the microbiome profiles of >200,000 patients with urinary tract or wound infections who were
treated with known antibiotics before and after their infections, ML can be used to predict the risk of
treatment-induced gain of resistance on a patient-specific level.
● Indeed, urinary tract infection (UTI) patients treated with antibiotics that the ML system would not have
recommended resulted in significantly resistance (E). Both UTI (F) and wound infection (G) patients would suffer
far fewer reinfections if they’d have been prescribed antibiotics according to the ML system.
#stateofai | 44

● Very few biological samples can typically
be identiﬁed from reference libraries.
● Property-prediction transformers
outperform at predicting a suite of
medicinally-relevant chemical properties
like solubility, drug likeness, and synthetic
accessibility directly from MS/MS, without
using structure prediction intermediates or
reference lookups.
stateof.ai 2022
Interpreting small molecule mass spectra using transformers
Tandem mass spectrometry (MS/MS) is commonly used in metabolomics, the study of small molecules in
biological samples. Less than 10% of small molecules can be identiﬁed from spectral reference libraries as most
of nature’s chemical space is unknown.Transformers enable fast, accurate, in silico, characterization of the
molecules in metabolic mixtures, enabling biomarker and natural product drug discovery at scale.
#stateofai | 45

● The researchers had trained their “MegaSyn” model to maximize bioactivity
and minimize toxicity. To design toxic molecules, they kept the same model,
but now simply training it to maximize both bioactivity and toxicity. They used
a public database of drug-like molecules.
● They directed the model towards generation of the nerve agent VX, known to
be one of the most toxic chemical warfare agents.
● However, as is the case with regular drug discovery, ﬁnding molecules with a
high predicted toxicity doesn’t mean it is easy to make them. But as drug
discovery with AI in the loop is being dramatically improved, we can imagine
best practices in drug discovery diffusing into building cheap biochemical
weapons.
Researchers from Collaborations Pharmaceuticals and King’s College London showed that machine learning
models designed for therapeutic use can be easily repurposed to generate biochemical weapons.
stateof.ai 2022
Drug discovery, the ﬂagship “AI for good” application, is not immune to misuse
#stateofai | 46

stateof.ai 2022
Compared to US AI research, Chinese papers focus more on surveillance related-tasks.
These include autonomy, object detection, tracking, scene understanding, action and
speaker recognition.
#stateofai | 47
Comparing data modalities in Chinese vs. US papers Comparing machine learning tasks in Chinese vs. US papers
Red = more common in China
Blue = more common in the US

stateof.ai 2022
While US-based authors published more AI papers than Chinese peers in 2022, China and
Chinese institutions are growing their output at a faster rate
#stateofai | 48
# papers published in 2022 and change vs. 2021
+11%
+24%
-2%
+3%
+10%
+4%
# papers published in 2022 and change vs. 2021
+27%
+13%
+13%
+11%
+1%

stateof.ai 2022
Chinese institutions author 4.5x the number of papers than American institutions since 2010.
The China-US AI research paper gap explodes if we include the Chinese-language
database, China National Knowledge Infrastructure
#stateofai | 49

Section 2: Industry
stateof.ai 2022
#stateofai | 50

stateof.ai 2022
NVIDIA’s FY 2021 datacenter revenue came in at $10.6B. In Q4 2021, they recognised $3.26B, which on an
annualised basis is greater than the combined valuation of top-3 AI semiconductor startups. NVIDIA has over 3
million developers on their platform and the company’s latest H100 chip generation is expected to deliver 9x
training performance vs. the A100. Meanwhile, revenue ﬁgures for Cerebras, SambaNova and Graphcore are not
publicly available.
#stateofai | 51
Do upstart AI chip companies still have a chance vs. NVIDIA’s GPU?
$13 billion
Latest private valuation
$4 billion
$2.8 billion
$5.1 billion
Annualised datacenter revenue

stateof.ai 2022
GPUs are 131x more commonly used than ASICs, 90x more than chips from Graphcore, Habana, Cerebras,
SambaNova and Cambricon combined, 78x more than Google’s TPU, and 23x more than FPGAs.
#stateofai | 52
NVIDIA’s chips are the most popular in AI research papers…and by a massive margin
23x
78-131x
log scale

stateof.ai 2022
The V100, released in 2017, is NVIDIA’s workhorse chip, followed by the A100 that was released in 2020.The
H100 is hotly awaited in 2022. Of the major AI chip challengers, Graphcore is cited most often.
#stateofai | 53
For NVIDIA, the V100 is most popular, and Graphcore is most used amongst challengers
Number of AI papers citing use of speciﬁc NVIDIA cards Number of AI papers citing use of speciﬁc AI chip startups

Announced at $40B, NVIDIA’s attempted acquisition of Arm fell through due to signiﬁcant geopolitical and anti
competition pushback. Nonetheless, NVIDIA’s enterprise value grew by $295B during the period (!!)
stateof.ai 2022
#stateofai | 54
NVIDIA fails to acquire Arm and grows its revenue 2.5x and valuation 2x during the deal
Deal announced Deal cancelled

NVIDIA has been investing heavily in AI research and producing some of the best works in imaging over the years.
For instance, their latest work on view synthesis just won the best paper award at SIGGRAPH, one of the most
prestigious computer graphics conferences. But NVIDIA has now gone a step further and applied their
reinforcement learning work to design their next-generation AI chip, the H100 GPU.
stateof.ai 2022
#stateofai | 55
NVIDIA reaps rewards from investing in AI research tying up hardware and software

stateof.ai 2022
The hyperscalers and challenger AI compute providers are tallying up major AI compute partnerships, notably
Microsoft’s $1B investment into OpenAI. We expect more to come.
#stateofai | 56
David teaming up with Goliath: training large models requires compute partnerships
None yet?
None yet?

“We think the most beneﬁts will go to whoever has the biggest computer”– Greg Brockman, OpenAI CTO
#stateofai | 57
In a gold rush for compute, companies build bigger than national supercomputers
Current NVIDIA A100 GPU count Future NVIDIA H100 GPU count
*
e
s
t
i
m
a
t
e
d
stateof.ai 2022

stateof.ai 2022
In 1962 the US government bought all integrated circuits in the world, supercharging the development of this
technology and its end markets. Some governments are providing that opportunity again, as “buyers of ﬁrst
resort” for AI companies. With access to unique high-quality data, companies could gain an edge in building
consumer or enterprise AI software.
The compounding effects of government contracting in AI
#stateofai | 58
● Researchers examined Chinese facial recognition AI companies and
showed a causal relationship between the number of government
contracts they signed and the cumulative amount of general AI
software they produced. Unsurprisingly, leadership in the computer
vision space has largely been ceded to Chinese companies now.
● The principle should stand in other heavily regulated sectors, like
defence or healthcare, which build an expertise through unique data
that is transferable to everyday AI products.

stateof.ai 2022
Meta’s release of the BlenderBot3 chatbot for free public use in August 2022 was faced with catastrophic press
because the chatbot was spitting misinformation. Meanwhile, Google, which published a paper on their chatbot
LaMDA in May 2021, had decided to keep the system in-house. But a few weeks after BlenderBot’s release,Google
announced a larger initiative called “AI test kitchen”, where regular users will be able to interact with Google’s
latest AI agents, including LaMDA.
● Large-scale release of AI systems to the 1B+ users of Google and
Facebook all but ensures that every ethics or safety issue with these
systems will be surfaced, either by coincidence or by adversarially
querying them. But only by making these systems widely available can
these companies ﬁx those issues, understand user behaviour and create
useful and proﬁtable systems.
● Running away from this dilemma, 4 of the authors of the paper
introducing LaMDA went on to found/join Character.AI, which describes
itself as “an AI company creating revolutionary open-ended conversational
applications”. Watch this space…
#stateofai | 59
How should big tech deal with their language model consumer products?

Once considered untouchable, talent from Tier 1 AI labs is breaking loose and becoming entrepreneurial.
Alums are working on AGI,AI safety, biotech, ﬁntech, energy, dev tools and robotics. Others, such as Meta, are
folding their centralised AI research group after letting it run free from product roadmap pressure for almost
10 years. Meta concluded that “while the centralized nature of the [AI] organization gave us leverage in some
areas it also made it a challenge to integrate as deeply as we would hope.”
stateof.ai 2022
#stateofai | 60
DeepMind and OpenAI alums form new startups and Meta disbands its core AI group

stateof.ai 2022
All but one author of the landmark paper that introduced transformer-based neural networks have left Google
to build their own startups in AGI, conversational agents,AI-ﬁrst biotech and blockchain.
#stateofai | 61
Attention is all you need… to build your AI startup
$580M
$225M
$125M
$65M
Amount raised in 2022

OpenAI’s Codex quickly evolved from research (July 2021) to open commercialization (June 2022) with
(Microsoft’s) GitHub Copilot now publicly available for $10/month or $100/year.Amazon followed suit by
announcing CodeWhisperer in preview in June 2022. Google revealed that it was using an internal ML-powered
code completion tool (so maybe in a few years in a browser IDE?). Meanwhile, with its 1M+ users, tabnine
raised $15M, promising accurate multiline code completions.
stateof.ai 2022
#stateofai | 62
AI coding assistants are deployed fast, with early signs of developer productivity gains
Metrics for Google’s coding assistant. Users are 10k+
Google-internal developers (5k+ for multi-line
experiments).
Single line Multi-line
Fraction of code added by ML 2.6% 0.6%
Average characters per accept 21 73
Acceptance rate (for suggestions visible for >750ms) 25% 34%
Reduction in coding iteration duration 6% -

And many more assets in early discovery stages. We expect early clinical trial readouts from 2023 onwards.
stateof.ai 2022
#stateofai | 63
AI-ﬁrst drug discovery companies have 18 assets in clinical trials, up from 0 in 2020
# of assets per pipeline stage per company % of assets per pipeline stage overall
Updated as of 26 Aug 2022

A study of 6,151 successful phase transitions between 2011–2020 found that it takes 10.5 years on average
for a drug to achieve regulatory approval. This includes 2.3 years at Phase I, 3.6 years at Phase II, 3.3 years at
Phase III, and 1.3 years at the regulatory stage. What’s more, it costs $6.5k on average to recruit one patient
into a clinical trial. With 30% of patients eventually dropping out due to non-compliance, the fully-loaded
recruitment cost is closer to $19.5k/patient. While AI promises better drugs faster, we need to solve for the
physical bottlenecks of clinical trials today.
stateof.ai 2022
#stateofai | 64
Can AI and compute bend the physical reality of clinical trial chokepoints?
# of registered studies (ClinicalTrials.gov EOY) Stepwise probability of drug success

● A large pre-trained protein language model was
trained on viral spike protein sequences of variants.
● New spike protein variants are fed to a transformer
that outputs embeddings and a probability
distribution of the 20 natural amino acids for each
position to determine how this would affect immune
escape and ﬁtness.
● The red dash line indicates the date when the EWS
predicted the variant would be high-risk and the
green dash-dot line is when the WHO designated the
variant. In almost all cases, EWS alerted several
months before the WHO designation.
stateof.ai 2022
mRNA vaccine leader, BioNTech, and enterprise AI company, InstaDeep, collaboratively built and validated an
Early Warning System (EWS) to predict high-risk variants.The EWS could identify all 16 WHO-designated
variants on average more than one and a half months prior to ofﬁcially receiving the designation.
#stateofai | 65
Predicting the evolution of real-world covid variants using language models

● Due to a shortage of radiologists and an increasing volume of
imaging, the diagnostic task of assessing which X-rays contain
disease and which don’t is challenging.
● Oxipit’s ChestLink is a computer vision system that is tasked
with identifying scans that are normal.
● The system is trained on over a million diverse images. In a
retrospective study of 10,000 chest X-rays of Finnish primary
stateof.ai 2022
Lithuanian startup Oxipit received the industry’s first autonomous certification for their computer vision-based
diagnostic.The system autonomously reports on chest X-rays that feature no abnormalities, removing the need
for radiologists to look at them.
#stateofai | 66
The first regulatory approval for an autonomous AI-first medical imaging diagnostic
health care patients, the AI achieved a sensitivity of 99.8% and specificity of 36.4 % for recognising clinically
significant pathology on a chest X-ray.
● As such, the AI could reliably remove 36.4% of normal chest X-rays from a primary health care population data
set with a minimal number of false negatives, leading to effectively no compromise on patient safety and a
potential significant reduction of workload.

Universities are a hotbed for AI spinouts: the UK case study
stateof.ai 2022
Universities are an important source of AI companies including Databricks, Snorkel, SambaNova, Exscientia and
more. In the UK, 4.3% of UK AI companies are university spinouts, compared to 0.03% for all UK companies.AI is
indeed among the most represented sectors for spinouts formation. But this comes at a steep price: Technology
Transfer Ofﬁces (TTOs) often negotiate spinout deal terms which are unfavourable to founders, e.g. a high equity
share in the company or royalties on sales.
#stateofai | 67

Spinout.fyi: an open database to help founders and policymakers ﬁx the spinout problem
stateof.ai 2022
Spinout.fyi crowdsourced a database of spinout deal terms from founders representing >70 universities all over
the world.The database spans AI and non-AI companies across different product categories (software, hardware,
medical, materials, etc.), and shows that the UK situation, while particularly discouraging for founders, isn’t
isolated. Only a few regions stand out as being founder-friendly, like the Nordics and Switzerland (ETH Zürich in
particular).A major reason for the current situation is the information asymmetry between founders and TTOs,
and the spinout.fyi database aims to give founders a leg up in the process.
#stateofai | 68

As 5-year programmes in Berkeley and Stanford wrap up, what comes next?
stateof.ai 2022
In 2011, UC Berkeley launched the “Algorithms, Machines, and People” (AMPLab) as a 5-year collaborative
research agenda amongst professors and students, supported by research agencies and companies.The program
famously developed the critical Big Data technology Spark (spun out as Databricks), as well as Mesos (spun out as
Mesosphere).This hugely successful program was followed in 2017 by the “Real-time intelligence secure
explainable systems” (RISELab) at Berkeley and “Data Analytics for What’s Next” (DAWN) at Stanford, which
focused on AI technologies. RISELab created the Ray ML workload manager (spun out as Anyscale), and DAWN
created and spun out the Snorkel active labelling platform. Will other universities and countries learn from the
successes of the 5-year model to fund ambitious open-source research with high spinout potential?
#stateofai | 69
Lab name OSS project created Spinouts that emerged
$38B val $250M raised
$1B val
$1B val
2011-16
2017-22

In 2022, investment in startups using AI has slowed down along with the broader market
stateof.ai 2022
Private companies using AI are expected to raise 36% less money in 2022* vs. last year, but are still on track to
exceed the 2020 level.This is comparable with the investment in all startups & scaleups worldwide.
$100B
$50B
$150B
0
▊ $250M+ ▊ $100-250M ▊ $40-100M (Series C) ▊ $15-40M (Series B)
▊ $4-15M (Series A) ▊ $1-4M (Seed) ▊ $0-1M (Pre-Seed)
$111.4B
$70.9B*
2
0
2
1
2
0
2
0
2
0
1
9
2
0
1
8
2
0
1
7
2
0
1
6
2
0
1
5
2
0
1
4
2
0
1
3
2
0
1
2
2
0
1
1
2
0
1
0
2
0
2
2
Y
T
D
Worldwide investment in startups & scaleups
using AI by round size » view online
$600B
$400B
$800B
0
▊ $250M+ ▊ $100-250M ▊ $40-100M (Series C) ▊ $15-40M (Series B)
▊ $4-15M (Series A) ▊ $1-4M (Seed) ▊ $0-1M (Pre-Seed)
2
0
2
1
2
0
2
0
2
0
1
9
2
0
1
8
2
0
1
7
2
0
1
6
2
0
1
5
2
0
1
4
2
0
1
3
2
0
1
2
2
0
1
1
2
0
1
0
2
0
2
2
Y
T
D
Worldwide investment in all startups & scaleups
by round size » view online
$726.5B
$399.7B
$553.2B*
$47.5B
$200B
$351.7B
$69.3B
-36%
-24%
In this slide, startups & scaleups using AI include both startups & scaleups with AI-ﬁrst and AI-enabled products and solutions.
(*) estimated amount to be raised by the end of 2022

The drop in investment is most noticeable in megarounds
stateof.ai 2022
The drop in VC investment is most noticeable in 100M+ rounds, whereas smaller rounds are expected to amount
to $30.9B worldwide by the end of 2022, which is almost on track with the 2021 level.
$40B
$20B
$80B
0
▊ $250M+ ▊ $100-250M
2
0
2
1
2
0
2
0
2
0
1
9
2
0
1
8
2
0
1
7
2
0
1
6
2
0
1
5
2
0
1
4
2
0
1
3
2
0
1
2
2
0
1
1
2
0
1
0
2
0
2
2
Y
T
D
Worldwide investment in startups & scaleups
using AI by round size » view online
$30B
$20B
$40B
0
▊ $40-100M (Series C) ▊ $15-40M (Series B) ▊ $4-15M (Series A)
▊ $1-4M (Seed) ▊ $0-1M (Pre-Seed)
2
0
2
1
2
0
2
0
2
0
1
9
2
0
1
8
2
0
1
7
2
0
1
6
2
0
1
5
2
0
1
4
2
0
1
3
2
0
1
2
2
0
1
1
2
0
1
0
2
0
2
2
Y
T
D
Worldwide investment in startups & scaleups using
AI by round size » view online
$10B
$77.5B
$34.9B*
$25.2B
$44.4B
$60B
$33.9B
$24.9B
$22.3B
$30.9B*
-55%
-9%

Public valuations have dropped in 2022, while private keep growing
stateof.ai 2022
Combined public enterprise value (EV) has dropped to the 2020 level. Meantime, private valuations keep growing,
with the combined EV already reaching $2.2T, up 16% from last year.
$4.0T
$2.0T
$8.0T
0
2
0
2
1
2
0
2
0
2
0
1
9
2
0
1
8
2
0
1
7
2
0
1
6
2
0
1
5
2
0
1
4
2
0
1
3
2
0
1
2
2
0
1
1
2
0
1
0
2
0
2
2
Y
T
D
Combined EV of public startups & scaleups using
AI by launch year; worldwide » view online
$1.5T
$1.0T
0
2
0
2
1
2
0
2
0
2
0
1
9
2
0
1
8
2
0
1
7
2
0
1
6
2
0
1
5
2
0
1
4
2
0
1
3
2
0
1
2
2
0
1
1
2
0
1
0
2
0
2
2
Y
T
D
Combined EV of privately owned startups & scaleups
using AI by launch year; worldwide » view online
$500B
$6.0T
$2.0T
$2.5T
$10.0T $9.6T
$6.8T $6.8T
$2.2T
$1.4T
$1.9T
▊ 2015-2022 YTD ▊ 2010-2014 ▊ 2005-2009 ▊ 2000-2004
▊ 1995-1999 ▊ 1990-1994
▊ 2015-2022 YTD ▊ 2010-2014 ▊ 2005-2009 ▊ 2000-2004
▊ 1995-1999 ▊ 1990-1994
-29%
+16%

The US leads by the number of AI unicorns, followed by China & the UK
stateof.ai 2021
Number of AI unicorns Combined enterprise value (2022 YTD) Examples
United States
China
United Kingdom
Israel
Germany
Canada
Singapore
Switzerland
292
69
24
14
10
7
6
6
$4.6T
$1.4T
$207B
$53B
$56B
$12B
$39B
$14B
Countries with the largest number of AI unicorns » view online
The US has created 292 AI unicorns, with the combined enterprise value of $4.6T.
#stateofai | 73

Investment in the USA accounts for more than half of the worldwide VC
stateof.ai 2022
2019 2020 2021
2016 2017 2018
2015
2014
2012 2013
2011
2010
▊ China
▊ United Kingdom
▊ Rest of the world
$60B
$40B
$140B
$20B
0
Despite signiﬁcant drop in investment in US-based startups & scaleups using AI, they still account for more than
half of the AI investment worldwide.
2022
YTD
$80B
$100B
$120B
▊ EU-27, Switzerland & Norway
▊ USA
$25.1B
$63.6B
53% in 2022 YTD
vs. 57% in 2021
$8.9B
$7.5B
$3.8B
$4.3B
$7.5B
$13.1B
Amount invested in the companies using AI

Enterprise software is the most invested category globally, while robotics captures the largest
share of VC investment into AI
stateof.ai 2022
Industries
Amount invested in
startups & scaleups
using AI, 2010-2022 YTD
Amount invested in
startups & scaleups
using AI, 2021-2022 YTD
Amount invested in startups
& scaleups using AI as %
of all VC, 2021-2022 YTD
Investment in startups & scaleups
using AI, 2021-2022 YTD;
number of rounds
$6.5B
$6.0B
$5.2B
$4.9B
$3.4B
$2.5B
$2.4B
$2.2B
$2.1B
$927M
$292M
$85M
$61.7B
$48.9B
$44.1B
$43.0B
$37.9B
$27.4B
$17.9B
$17.8B
$14.7B
$13.1B
$10.6B
$7.7B
$133.7B
$109.5B
$629M
$1.5B
$2.4B
$568M
$421M
$28M
$349M
$408M
$348M
$33M
$69M
$5M
$21.9B
$18.4B
$6.0B
$13.5B
$12.2B
$4.0B
$2.7B
$6.8B
$2.5B
$3.6B
$3.4B
$617M
$44.7B
$22.8B
$25.1B
3%
8%
14%
7%
4%
1%
2%
9%
3%
1%
2%
4%
14%
71%
10%
24%
25%
13%
13%
41%
9%
6%
24%
2%
30%
21%
13%
55
31
103
88
56
22
48
16
56
14
13
4
689
367
220
406
306
182
44
73
143
231
27
148
1288
332
660
$87.8B
Gaming
Home living
Jobs recruitment
Legal
Sports
Music
Fashion
Hosting
Wellness beauty
Event tech
Kids
Dating
Health
Robotics
Food
Marketing
Security
Media
Telecom
Semiconductors
Education
Energy
Travel
Real estate
Enterprise software
Transportation
Fintech
1%
1%
3%
2%
1%
1%
1%
<1%
1%
1%
<1%
<1%
3%
8%
1%
4%
5%
2%
2%
2%
3%
2%
1%
1%
8%
2%
3%
% of AI startups

stateof.ai 2022
Acquisitions are on track to exceed the 2021 level
300
200
400
100
0
500
While the number of IPOs and SPAC IPOs declined sharply, the number of acquisitions is on track to exceed the
2021 level
2021
2020
2019
2018
2017
2016
2015
2014
2013
2012
2011
2010
2022
YTD
▊ Acquisition ▊ Buyout ▊ IPO ▊SPAC IPO
Number of exits among the companies using AI;
worldwide » view online
Number of exits in 2022 among the companies
using AI; worldwide » view online
Technology, data & analytics
for real estate
$13.1B Acquisition
May 2022
SaaS for planning
& forecasting
$10.7B Buyout
Mar 2022
Customer experience
tools
$10.2B Buyout
Jun 2022
Tax compliance
software
$8.4B Buyout
Aug 2022
Consumer robotics
company
$1.7B Acquisition
Aug 2022
Autonomous, continuous
workload optimization
$650M Acquisition
Mar 2022
Mobile device & application
security solutions
$525M Buyout
Mar 2022
Cloud-native SOAR
platform
$500M Acquisition
Jan 2022
Developing autonomous
mobility platform
$469 Acquisition
Aug 2022
436
285
253
211
332
259

stateof.ai 2022
2019 2020 2021
2016 2017 2018
2015
2014
2012 2013
2011
2010
▊ China
▊ United Kingdom
▊ Rest of the world
150
100
500
50
0
2022
YTD
200
250
300
▊ EU-27, Switzerland & Norway
▊ USA
350
400
450
108
222
94
97
25
1
33
9
Number of exits among the companies using AI
The number of exits in EU-27, Switzerland & Norway has already exceeded 2021 levels
With 108 exits to date, the US hasn’t yet reached half of the 2021 level, while the EU, Switzerland & Norway
combined have already exceeded the 2021 number.

Investment in SaaS startups & scaleups using AI is expected to reach $41.5B by the end
of the year, down 33% from last year, but higher than in 2020
stateof.ai 2021
2019 2020 2021
2016 2017 2018
2015
2014
2012 2013
2011
2010
$20.0B
$80.0B
0
2022
YTD
$40.0B
108
222
VC investment in AI SaaS startups & scaleups » view online
▊ $250M+
▊ $100-250M
▊ $40-100M (Series C)
▊ $15-40M (Series B)
▊ $4-15M (Series A)
▊ $1-4M (Seed)
▊ $0-1M (Pre-Seed)
$61.8B
$37.7B
$41.5B*
-33%
$31.1B
$60.0B

The combined EV of public and private SaaS startups & scaleups using AI now amounts to
$2.3T, down 26% from last year, but still higher than in 2020
stateof.ai 2021
2019 2020 2021
2016 2017 2018
2015
2014
2012 2013
2011
2010
$1.0T
$4.0T
0
2022 YTD
$2.0T
Combined EV of AI SaaS startups & scaleups by launch year globally »
view online
$3.0T
$3.1T
$2.1T
$2.3T
-26%
▊ 2015-2022 YTD
▊ 2010-2014
▊ 2005-2009
▊ 2000-2004
▊ 1995-1999
▊ 1990-1994

stateof.ai 2021
The combined EV of private SaaS startups & scaleups using AI keeps growing and has
already reached $1.1T, up 12% from last year
$500B
$1.0T
0
$1.5T
2021
2020
2019
2018
2017
2016
2015
2014
2013
2012
2011
2010
2022
YTD
Combined EV of privately owned AI SaaS startups & scaleups by
launch year; worldwide » view online
Top valued privately owned startups & scaleups
using AI » view online
Financial infrastructure
platform
USA
Valuation: $68.4B
Global payments solution
provider
United Kingdom
Valuation: $40B
Lakehouse platform to unify
data, analytics and AI
USA
Valuation: $38B
Autonomous driving
technology
USA
Valuation: $30B
Online education
platform
China
Valuation: $15.5B
Fintech-as-a-service
platform
United Kingdom
Valuation: $15B
Process mining
software
Germany
Valuation: $13B
Business cloud
software products
USA
Valuation: $13B
AI-powered writing
assistant
USA
Valuation: $13B
▊ 2015-2022 YTD ▊ 2010-2014 ▊ 2005-2009 ▊ 2000-2004
▊ 1995-1999 ▊ 1990-1994
$1.0T
$895B
$489B
+12%
#stateofai | 80

Section 3: Politics
stateof.ai 2022
#stateofai | 81

stateof.ai 2022
The compute requirements for large-scale AI experiments has increased >300,000x in the last decade. Over the
same period, the % of these projects run by academics has plummeted from ～60% to almost 0%. If the AI
community is to continue scaling models, this chasm of “have” and “have nots” creates signiﬁcant challenges for AI
safety, pursuing diverse ideas, talent concentration, and more.
A widening compute chasm is separating industry from academia in large model AI
#stateofai | 82

There is a growing appreciation that AI is an engineering science in which the objects of study need to ﬁrst be
built. Western academics and governments are starting to wake up to this reality, most notably through the
National AI Research Resource process in the US. While spending years on consultations and marketing however,
others in China and outside academia are ﬁnding creative ways to do large-scale AI projects.
stateof.ai 2022
Slow progress in providing academics with more compute leaves others to act faster
#stateofai | 83
National AI
Initiative Act
enacted
Jan 2021
Stanford opens Center for Research on
Foundation Models
Research Collectives
Governments / Academia
NAIRRTF Final
Report
Dec 2022
May
Dec Apr
Feb 2022 Jul
Oct
Aug
The Pile
(Eleuther)
July Sep Oct
NAIRRTF
Meeting 5
NAIRRTF
Meeting 6
NAIRRTF
Meeting 7
NAIRRTF
Meeting 8
NAIRRTF
Meeting 9
NAIRRTF
Meeting 4
National AI
Research Resource
Task Force Created
GPT-J-6B
(Eleuther)
GPT-NeoX-20B
(Eleuther)
Aug
Stable Diffusion
(Stability)
GLM-130B
(Tsinghua)
GPT-Neo
(Eleuther)
Mar
Jun
NAIRRTF
Meeting 2
NAIRRTF
Meeting 3
NAIRRTF
Meeting 1
BLOOM-178B
(BigScience)
NAIRRTF
Meeting 10

● The most notable large-scale academic project this year
came from China: Tsinghua’s GLM-130B LLM.
● Eleuther, the original AI research collective, released the
20B parameter GPT-NeoX. However, core members have
since moved on to OpenAI, Stability and Conjecture.
● Hugging Face led the BigScience initiative, releasing the
178B parameter BLOOM multilingual LLM.
● Stability came out of nowhere, obtained 4,000 A100
GPUs, brought together multiple open-source
communities and created Stable Diffusion.
Decentralized research projects are gaining members, funding and momentum.They are succeeding at ambitious
large-scale model and data projects that were previously thought to be only possible in large centralised
technology companies – most visibly demonstrated by the public release of Stable Diffusion.
stateof.ai 2022
The baton is passing from academia to decentralized research collectives
#stateofai | 84
Large-Scale AI Results
Graph data source: Sevilla et al. Parameter, Compute and Data Trends in Machine Learning

● Stability has embedded itself as a compute platform for
independent and academic open-source AI communities:
supporting LAION for building a dataset of 5B image-text pairs
and training an open-source CLIP model, and supporting the
CompVis group’s research in efﬁcient diffusion models. It funds
PhD students to work on community projects, and has directly
hired generative AI artists, core members of Eleuther, and
renowned ML researchers such as David Ha.
● Stable Diffusion cost <$600K to train, and while weights were
released, access is also sold through the DreamStudio API.
Where there was previously a dependence on ad-hoc compute
donations to enable large-scale projects, Stability is pioneering a
new approach of structured compute and resource provision for
open-source communities, while also commercializing these
projects with revenue-sharing for developers.
stateof.ai 2022
Stability AI is attempting a new paradigm in commercializable open-source AI
#stateofai | 85

● Epirus, founded in 2018, has built a next-generation electromagnetic pulse
weapon capable of defeating swarms of drones that pose threats to human
safety. Sweden’s Saab is also making efforts towards AI-driven automation of
electronic warfare: they built the COMINT and C-E.SM sensors to balance
automated and operator-controlled surveillance depending on the context on the
ﬁeld. The company is also collaborating with defense startup, Helsing.
● Modern Intelligence, founded in 2020, builds a platform-independent AI for
geospatial sensor data fusion, situational awareness and maritime surveillance.
● Meanwhile, through both organic and inorganic growth, Anduril has expanded its
autonomous hardware platforms. For example, Anduril acquired Area-I to launch
a new product in Air Launched Effects with an increased payload, data sharing
and integration capabilities with other UAVs. Anduril also expanded into
Underwater Autonomous Vehicles by acquiring Dive Technologies.
Defense technology companies are applying AI to electronic warfare, geospatial sensor fusion, and to create
autonomous hardware platforms.
stateof.ai 2022
AI continues to be infused into a greater number of defense product categories
#stateofai | 86

● Nato published their AI Strategy and announced a $1B fund to invest
in companies working on a range of dual-use technologies. It was
described as the world’s ﬁrst ‘multi-sovereign venture capital fund’
spanning 22 nations.
● Helsing, a European defense AI company, announced a €102.5M
Series A led by Daniel Ek of Spotify.
● Microsoft, Amazon and Google continue to compete for a major role
in defense - most notably Microsoft’s $10B contract with the
Pentagon was cancelled after a lawsuit from Amazon. The new
beneﬁciaries of the contract will now be announced in late 2022.
● Anduril landed their largest DoD contract to date and is now
reportedly valued at $7B.
● Shield AI, developer of military drones, raised at a $2.3B valuation
Heavily funded start-ups and Amazon, Microsoft, and Google continue to normalise the use of AI in Defense.
stateof.ai 2022
AI in defense gathers big funding momentum
#stateofai | 87

● GIS Arta is a homegrown application developed prior to Russia’s
invasion based on lessons learned from the conﬂict in the
Donbas.
● It’s a guidance command and control system for drone, artillery
or mortar strikes.
● The app ingests various forms of intelligence (from drones, GPS,
forward observers etc) and converts it into dispatch requests for
reconnaissance and artillery.
● GIS Arta was allegedly developed by a volunteer team of
software developers led by Yaroslav Sherstyvk, inspired by the
Uber taxi model.
The use of geospatial (GIS) software has reportedly reduced the decision chain around artillery from 20 minutes
to under one minute.
stateof.ai 2022
Ukraine’s homegrown geospatial intelligence GIS Arta software is a sign of things to come
#stateofai | 88

Between 1990 and 2020, China accelerated its output of greenﬁeld fab projects by almost 7x while the US slowed
down by 2.5x. Moreover, while China and Taiwan fabs take roughly 650 days from construction start to being
production-ready, the US builds fabs 42% slower today than they did 30 years ago.
stateof.ai 2022
The Great Reshoring will be slow: US lags in new fab projects, which take years to build
#stateofai | 89
Total # of greenﬁeld fab projects Avg. # of days from build start to production

stateof.ai 2022
#stateofai | 90
● The bill poses a dilemma for Korean (e.g. Samsung), Taiwanese (e.g.
TSMC) and other manufacturers: if they accept US subsidies, then they
must pivot away from China without backlash from Beijing, which is
opposed to this ��friendshoring”.
● Since passing the bill, Micron announced a $40B investment in
memory chip manufacturing to increase US market share from 2% to
10%. Qualcomm will expand its US semiconductor production by 50%
by 2027 and in partnership with GlobalFoundries the two will invest
$4.2B to expand the latter’s upstate New York facility.
● CSET estimates the US should focus on its manufacturing capabilities
in leading-edge, legacy logic and DRAM (right chart).
The US CHIPS and Science Act of 2022: $250B for US semiconductor R&D and production
The bipartisan legislation was signed into law in August 2022. It provides for $53B to boost US-based
semiconductor R&D, workforce and manufacturing, as well as a 25% investment tax credit for semiconductor
manufacturers’ capital expenses.In exchange, recipients must not upgrade or expand their existing operations
in China for 10 years, nor can they use funds for share buybacks or to issue dividends.

stateof.ai 2022
#stateofai | 91
● Earlier this year, CSET analysed 24 public contracts awarded by Chinese PLA units and state-owned defense
enterprises in 2020. They found that nearly all of the 97 AI chips in these purchase orders were designed by
NVIDIA, AMD, Intel and Microsemi. Domestic AI chip companies were not featured. Thus, American chips are
arming Chinese defense capabilities.
● Chinese semiconductor manufacturers have been already cut off from advanced lithography machines made
by ASML and related equipment from Lam Research and Applied Materials.
● It is unlikely that domestic AI chip companies (e.g. Biren) can ﬁll the void: leading-edge node manufacturing is
still only possible by TSMC in Taiwan and because domestic talent, software and technology is still years away
from NVIDIA. China will still accelerate its development.
The US cuts China off from NVIDIA and AMD chips…will this spur Chinese AI R&D?
NVIDIA GPUs are used by all major Chinese technology companies (Baidu,Tencent et al.) and universities
(Tsinghua, Chinese Academy of Sciences et al.). Washington ordered NVIDIA and AMD to stop exporting their
latest AI chips (e.g. NVIDIA A100 and H100, and AMD M100 and M200) to China as a means of curbing their
use in applications that threaten American national security via China.The companies will have to provide
statistics on previous shipments and customer lists. Not having access to state of the art AI chips could stall a
large swath of Chinese industry if domestic suppliers don’t step into the void and fast.

stateof.ai 2022
#stateofai | 92
● The AI Act moves through the EU legislative process. The European Parliament has worked over the summer
on a compromise text to address tabled amendments and opinions by the Parliament’s various committees.
The compromise text is scheduled to go through the various stages of the voting process at the European
Parliament by the end of 2022.
● The AI Act is expected to be voted into law in 2023, either under the Swedish or the Spanish Presidency of the
EU.
● Current realistic expectations are that the AI Act will become effective in the second-half of 2023.
The EU advances with its plans to regulate AI
In April 2021, the EU tabled a proposal for regulating the placement on the market and use in the EU of AI
systems (the “AI Act”).The proposal introduces certain minimum requirements (e.g. mainly information
obligations) that all AI systems in use must meet. It also introduces more elaborate requirements (e.g. risk
assessments, quality management) with respect to AI systems that pose higher risks to users.The AI Act bans
the use of certain types of AI-based techniques (e.g. social scoring, real-time biometric remote identiﬁcation
(subject to exceptions),“subliminal techniques”).
Source: Dessislava Fessenko

stateof.ai 2022
#stateofai | 93
The EU aims at quick operationalization of the AI Act
The EU aims at quick operationalization of the requirements under the AI Act through standardization, setting
up of testing facilities, and launch of pan-European and national regulatory sandboxes.
● European standardization efforts are already underway. The EU standardization organizations CEN and
CENELEC have already commenced preparatory works on standardization and expect to be requested to
develop relevant sets of standards by 31 October 2024.
● The EU appears to favor testing of high-risk AI systems, in either controlled or even possibly in real-world
conditions, as a suitable mode for supporting and promoting compliance with the AI Act among businesses of
all sizes.
● Pan-European and national regulatory sandboxes start to emerge in the EU. Spain launched the ﬁrst one in
June 2022. Other EU member states (e.g. the Czech Republic) have announced similar plans. Sandboxes are
considered by EU regulators as suitable testbeds for technical, policy and standardization solutions. They are
also intended as a medium for supporting small and medium-sized businesses in particular in attaining
compliance with the AI Act.
Source: Dessislava Fessenko

Section 4: Safety
stateof.ai 2022
#stateofai | 94

stateof.ai 2022
#stateofai | 95
While AI advances rapidly, the safety of highly-capable future systems remains unclear
While many concerns still appear speculative, early AI pioneers considered that highly capable and
economically integrated AI systems of the future could fail catastrophically and pose a risk to humanity,
including through the emergence of behaviours directly opposed to human oversight and control.
“... it seems probable
that once the
machine thinking
method has started,
it would not take
long to outstrip our
feeble powers. …
At some stage
therefore we should
have to expect the
machines to take
control”
“Thus the ﬁrst
ultraintelligent
machine is the last
invention that man
need ever make,
provided that the
machine is docile
enough to tell us
how to keep it under
control.”
“The problem is that,
with such powerful
machines, it would
require but the
slightest accident of
careless design for
them to place their
goals ahead of ours”
Alan Turing
1951
I.J. Good
1965
Marvin Minsky
1984

stateof.ai 2022
#stateofai | 96
The UK is taking the lead on acknowledging these uncertain but catastrophic risks
The UK’s national strategy for AI, published in late 2021, notably made multiple references to AI safety and the
long-term risks posed by misaligned AGI.
● “While the emergence of Artificial General Intelligence (AGI) may seem like a
science fiction concept, concern about AI safety and non-human-aligned systems is
by no means restricted to the fringes of the field.”
● “We take the firm stance that it is critical to watch the evolution of the technology,
to take seriously the possibility of AGI and ‘more general AI’, and to actively direct
the technology in a peaceful, human-aligned direction.”
● “The government takes the long term risk of non-aligned AGI,and the
unforeseeable changes that it would mean for the UK and the world,seriously.”
● “[We must] establish medium and long term horizon scanning functions to
increase government’s awareness of AI safety.”
● “[We must] work with national security, defence, and leading researchers to
understand how to anticipate and prevent catastrophic risks.”

● A survey of the ML community found that
69% believe AI safety should be prioritized
more than it currently is.
● A separate survey of the NLP community
found that a majority believe AGI is an
important concern we are making progress
towards. Over 70% believe AI will lead to
social change at the level of the Industrial
Revolution this century, and nearly 40%
believe AI could cause a catastrophe as bad
as nuclear war during that time.
Long dismissed as science ﬁction by mainstream AI research and academia, researchers are now shifting
consensus towards greater concern for the risks of human-level AI and superhuman AGI in the near future.
stateof.ai 2022
AI researchers increasingly believe that AI safety is a serious concern
#stateofai | 97
Note: The number in green represents the fraction of respondents who agree with the position out of all
those who took a side.The number in black shows the average predicted rate of agreement.

● New non-profit research labs include the Center for AI Safety and the Fund for
Alignment Research. The Centre for the Governance of AI was spun out as an
independent organization from the Future of Humanity Institute in Oxford.
● There was a huge increase in interest for education programmes with over 750
people taking part in the online AGI Safety Fundamentals course. New
scholarships were created, including the Vitalik Buterin PhD Fellowship in AI
Existential Safety.
● Notably, Ilya Sutskever, OpenAI’s Chief Scientist, has shifted to spending 50% of
his time on safety research.
Increased awareness of AI existential risk is leading to increased headcount,
with an estimated 300 researchers now working full-time on AI safety.
However, this is still orders of magnitude fewer researchers than are working in
the broader field, which itself is growing faster than ever (right chart).
stateof.ai 2022
AI safety is attracting more talent… yet remains extremely neglected
#stateofai | 98
Researchers by venue/field
30x

Increased awareness of AI existential risk has led to rapidly increasing funding for research into the safety of
highly-capable systems, primarily through donations and investments from sympathetic tech billionaires Dustin
Moskovitz (Open Philanthropy) and Sam Bankman-Fried (FTX Foundation). However, total VC and philanthropic
safety funding still trails behind resources for advanced capabilities research, not even matching DeepMind’s
2018 opex.
stateof.ai 2022
Funding secured, though trailing far behind what goes into capabilities
#stateofai | 99
Philanthropic AI Safety funding pales in comparison to AI capabilities funding*
*We include fundraises for Adept, Hugging Face, Cohere, AI21, Stability and Inﬂection under Capabilities VC and fundraises for Anthropic under Safety VC.

stateof.ai 2022
RLHF has emerged as a key method to finetune LLMs and align them with human values.This involves humans
ranking language model outputs sampled for a given input, using these rankings to learn a reward model of
human preferences, and then using this as a reward signal to finetune the language model with using RL.
● OpenAI started the year by finetuning GPT-3 using RLHF to
produce InstructGPT models that improved on helpfulness for
instruction-following tasks. Notably, the fine-tuning only
needed <2% of GPT-3’s pretraining compute, as well as 20,000
hours of human feedback. API users on average prefer these
models to the original ones.
● RLHF has also been used by both Anthropic and Deepmind to
improve the helpfulness, harmlessness, and correctness of
their language models. OpenAI has stated that RLHF will be a
core building block for it’s long-term approach to aligning
AGI.
#stateofai | 100
Language Model Alignment: Reinforcement Learning from Human Feedback (RLHF)

stateof.ai 2022
An RLHF preference model provides limited learning signal, compared to the full
expressiveness of language that humans use. NYU researchers demonstrated that
language models could be improved directly using human feedback written in
language.
● Notably, their method was highly data-efﬁcient, with only 100 samples of
human feedback they were able to ﬁnetune GPT-3 and improve it’s abilities on
a summarization task to human-level.
● On a synthetic task for removing offensive words from sentences, they observe
that only the largest GPT-3 models are capable of incorporating feedback
effectively, demonstrating another example of emergent behaviour at scale.
#stateofai | 101
Language Model Alignment: Reinforcement Learning from Human Feedback

stateof.ai 2022
As language models exhibit an increasing array of capabilities, it becomes
difficult to exhaustively evaluate their failure modes, inhibiting trust and
safe public deployment. DeepMind introduced automated “red teaming”, in
which manual testing can be complemented through using other language
models to automatically“attack” other language models to make them
exhibit unsafe behaviour, as determined by a separate classifier.
● Anthropic used manual red teaming to evaluate RLHF models, finding that
they are harder to attack and less harmful with increased model size.
#stateofai | 102
Language Model Alignment: Red Teaming
● In the future, a classifier could detect for
speculative risks such as power-seeking
behavior or malicious coding.

stateof.ai 2022
Interpreting deep neural networks when seen just as sequences of matrix operations is exceedingly difficult –
research in mechanistic interpretability instead seeks to reverse-engineer models into human-interpretable
computer programs, gaining an understanding of individual neurons as well as their collective behaviour.
● Researchers at Anthropic released significant analyses of small
transformer-based language models, focusing on a phenomenon of
“induction heads” that learn to copy and complete sequences which
have occurred before in a text. They find that these heads emerge
during “phase shifts” in training during which in-context learning
capabilities also emerge, and further developed a hypothesis that
these heads may be responsible for the majority of in-context
learning capabilities in large transformer models as well.
● Follow up work in this space has also brought to light ways in which
individual neurons become responsible for individual or multiple
semantic features, and ways to control this type of interpretability.
#stateofai | 103
Mechanistic interpretability – can we reverse-engineer neural networks?

State of AI Report 2022 - ONLINE.pdf

State of AI Report 2022 - ONLINE.pdf

Related slideshows

More Related Content

State of AI Report 2022 - ONLINE.pdf