Alan’s conservative countdown to AGI

Last update: Sep/2023

Get notified about this countdown’s progress, joining full subscribers from major government (USA, Singapore, Malta…), enterprise (DeepMind, Johnson & Johnson, Lockheed Martin…), and research institutions (RAND, HBS, Berkeley…). Get The Memo.

Definition of AGI


I use a slightly stricter definition for AGI that includes the ability to act on the physical world via embodiment. I appreciate that there were some approaches on getting to AGI that fully bypass embodiment or robotics.

Artificial general intelligence (AGI) is a machine capable of understanding the world as well as—or better than—any human, in practically every field, including the ability to interact with the world via physical embodiment.

And the short version: ‘AGI is a machine which is as good or better than a human in every aspect’.

The world and acceptance of AGI

Milestones & justifications (most recent at top)


Date Summary Links
Sep/2023 55%: OpenAI Gobi/GPT-5 leaks and analysis, early rumors from Sep/2023. Shared Google Doc
Sep/2023 55%: Harvard studies BCG consultants with GPT-4, ‘Consultants using [GPT-4] AI were significantly more productive (they completed 12.2% more tasks on average, and completed tasks 25.1% more quickly), and produced significantly higher quality results (more than 40% higher quality…)’ Paper (SSRN)
Sep/2023 55%: Google OPRO self-improves, ‘prompts optimized by OPRO outperform human-designed prompts by up to 8% on GSM8K [maths], and by up to 50% on Big-Bench Hard [IQ] tasks.’ Paper (arxiv)
Aug/2023 54%: GPT-4 scores in 99th percentile for Torrance Tests of Creative Thinking (wiki), questions by Scholastic Testing Service confirmed private/not part of training dataset. Article
Jul/2023 54%: Google DeepMind Robotics Transformer RT-2 (3x improvement over RT-1, 2x improvement on unseen scenarios to 62% avg. Progress towards Woz’s AGI coffee test.) Project page
Jul/2023 52%: Anthropic Claude 2: More HHH (TruthfulQA Claude 2=0.69 vs GPT-4=0.60) Anthropic (PDF)
Jul/2023 51%: Google DeepMind/Princeton: Robots that ask for help (‘modeling uncertainty that can complement and scale with the growing capabilities of foundation models.’) Project page
Jul/2023 51%: Microsoft LongNet: 1B token sequence length (‘opens up new possibilities for modeling very long sequences, e.g., treating a whole corpus or even the entire Internet as a sequence.’) Microsoft (arxiv)
Jun/2023 50%: Google DeepMind RoboCat (‘autonomous improvement loop…RoboCat not only shows signs of cross-task transfer, but also becomes more efficient at adapting to new tasks.’) DeepMind blog, Paper (PDF)
Jun/2023 50%: Microsoft introduces monitor-guided decoding (MGD) (‘improves the ability of an LM to… generate identifiers that match the ground truth… improves compilation rates and agreement with ground truth.’) Paper (arxiv)
Jun/2023 50%: Ex-OpenAI consultant uses GPT-4 for embodied AI in chemistry (‘instructions, to robot actions, to synthesized molecule.’) Paper (arxiv), notes
Jun/2023 50%: Harvard introduces ‘inference-time intervention’ (ITI) (‘At a high level, we first identify a sparse set of attention heads with high linear probing accuracy for truthfulness. Then, during inference, we shift activations along these truth-correlated directions. We repeat the same intervention autoregressively until the whole answer is generated.’) Harvard (arxiv)
Jun/2023 49%: Google DeepMind trains an LLM (DIDACT) on iterative code in their 86TB code repository (‘the trained model can be used in a variety of surprising ways… by chaining together multiple predictions to roll out longer activity trajectories… we started with a blank file and asked the model to successively predict what edits would come next until it had written a full code file. The astonishing part is that the model developed code in a step-by-step way that would seem natural to a developer’) Google Blog, Twitter
May/2023 49%: Ability Robotics combines an LLM with their humanlike android (robot), Digit. Agility Robotics (YouTube)
May/2023 49%: PaLM 2 breaks 90% mark for WinoGrande. For the first time, a large language model has breached the 90% mark on WinoGrande, a ‘more challenging, adversarial’ version of Winograd, designed to be very difficult for AI. Fine-tuned PaLM 2 scored 90.9%; humans are at 94%. PaLM 2 paper (PDF, Google)
May/2023 49%: Robot + text-davinci-003 (‘…we show that LLMs can be directly used off-the-shelf to achieve generalization in robotics, leveraging the powerful summarization capabilities they have learned from vast amounts of text data.’). Princeton/Google/others
Apr/2023 48%: Boston Dynamics + ChatGPT (‘We integrated ChatGPT with our [Boston Dynamics Spot] robots.’). Levatas
Mar/2023 48%: Microsoft introduces (‘We illustrate how TaskMatrix.AI can perform tasks in the physical world by [LLMs] interacting with robots and IoT devices… All these cases have been implemented in practice… understand the environment with camera API, and transform user instructions to action APIs provided by robots… facilitate the handling of physical work with the assistance of robots and the construction of smart homes by connecting IoT devices…’). Microsoft (arxiv)
Mar/2023 48%: OpenAI introduces GPT-4, Microsoft research on record that GPT-4 is ‘early AGI’ (‘Given the breadth and depth of GPT-4’s capabilities, we believe that it could reasonably be viewed as an early (yet still incomplete) version of an artificial general intelligence (AGI) system.’).
Microsoft’s deleted original title of the paper was ‘First Contact With an AGI System’.
Note that LLMs are still not embodied, and this countdown requires physical embodiment to get to 60%.
Microsoft Research
Mar/2023 42%: Google introduces PaLM-E 562B (PaLM-Embodied. ‘PaLM-E can successfully plan over multiple stages based on visual and language input… successfully plan a long-horizon task…’). Google
Feb/2023 41%: Microsoft used ChatGPT in robots, it self-improved (‘we were impressed by ChatGPT’s ability to make localized code improvements using only language feedback.’). Microsoft
Dec/2022 39%: Anthropic RL-CAI 52B trained by Reinforcement Learning from AI Feedback (RLAIF) (‘we have moved further away from reliance on human supervision, and closer to the possibility of a self-supervised approach to alignment’)., Anthropic paper (PDF)
Jul/2022 39%: NVIDIA’s Hopper (H100) circuits designed by AI (‘The latest NVIDIA Hopper GPU architecture has nearly 13,000 instances of AI-designed circuits’)., NVIDIA
May/2022 39%: DeepMind Gato is the first generalist agent, that can “play Atari, caption images, chat, stack blocks with a real robot arm, and much more”. Watch Alan’s video about Gato.
Jun/2021 31% Google’s TPUv4 circuits designed by AI (‘allowing chip design to be performed by artificial agents with more experience than any human designer. Our method was used to design the next generation of Google’s artificial intelligence (AI) accelerators, and has the potential to save thousands of hours of human effort for each new generation. Finally, we believe that more powerful AI-designed hardware will fuel advances in AI, creating a symbiotic relationship between the two fields’)., Nature, Venturebeat
Nov/2020 30%: Connor Leahy, Co-founder of EleutherAI, re-creator of GPT-2, creator of GPT-J & GPT-NeoX-20B, said about OpenAI GPT-3: “I think GPT-3 is artificial general intelligence, AGI. I think GPT-3 is as intelligent as a human. And I think that it is probably more intelligent than a human in a restricted way… in many ways it is more purely intelligent than humans are. I think humans are approximating what GPT-3 is doing, not vice versa.” Watch the video (timecode)
Aug/2017 20%: Google Transformer leads to big changes for search, translation, and language models. Read the launch in plain English.

AGI dates predicted based on this table

Thanks to Dennis Xiloj. In Jun/2023, using the current milestones and percentages, GPT-4 says AGI by 18/Jul/2025…

Thanks to The Memo reader BeginningInfluence55 for this more conservative version using polynomial regression. In Jul/2023, using the current milestones and percentages, this method says 80% probability of hitting AGI by Aug/2025…

Next milestones

– Around 50%: HHH: Helpful, honest, harmless as articulated by Anthropic, with a focus on groundedness and truthfulness. Mustafa Suleyman is the Co-founder of Google DeepMind, and Founder of Inflection AI (, and says: ‘LLM hallucinations will be largely eliminated by 2025’.

– Around 60%: Physical embodiment backed by a large language model. The AI is autonomous, and can move and manipulate. Current options include:

See related page: Humanoid robots ready for LLMs.

– Around 80%: Passes Steve Wozniak’s test of AGI: can walk into a strange house, navigate available tools, and make a cup of coffee from scratch (video with timecode).

Download source (PDF)
Permissions: Yes, you can use these visualizations anywhere, please leave the citation intact.

Where will AGI be born?

Related videos

Get The Memo

by Dr Alan D. Thompson · Be inside the lightning-fast AI revolution.
Thousands of paid subscribers. Readers from Microsoft, Tesla, Google AI...
Artificial intelligence that matters, as it happens, in plain English.
Get The Memo.

Dr Alan D. Thompson is an AI expert and consultant, advising Fortune 500s and governments on post-2020 large language models. His work on artificial intelligence has been featured at NYU, with Microsoft AI and Google AI teams, at the University of Oxford’s 2021 debate on AI Ethics, and in the Leta AI (GPT-3) experiments viewed more than 3.5 million times. A contributor to the fields of human intelligence and peak performance, he has held positions as chairman for Mensa International, consultant to GE and Warner Bros, and memberships with the IEEE and IET. He is open to consulting and advisory on major AI projects with intergovernmental organizations and enterprise.

This page last updated: 29/Sep/2023.