Alan's conservative countdown to AGI: 99%

AGI = artificial general intelligence = a machine that performs at the level of an average (median) human.

ASI = artificial superintelligence = a machine that performs at the level of an expert human in practically any field.

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’.

A reader asks ‘Does AGI really need physical embodiment? Think about Stephen Hawking. Was it a knock that he couldn’t move his body? Was that a part of his human intelligence? I’d argue that is a clear no. GPT-4 and Gemini are already a kind of proto ASI they are just glitchy and strange. The only thing they need to be ASI is more agency and automation that’s the final bottleneck. Clearly not ironed out yet sure. Putting them in a Boston dynamic robot won’t make them any smarter, the same way if we gave Hawking a magic cure to move again he would just be the same intelligence. Intelligence and embodiment are just not correlated at all with these systems; it wouldn’t change how smart they are, only utility, which is not the same thing.’

Here are some additional considerations for this thought experiment:

1. The definition of intelligence is not fully agreed upon, but may include ‘the ability to learn or understand or to deal with new or trying situations.’ Would it be possible to deal with a new or trying physical situation without embodiment?

2. Hawking had the benefit of more than two decades of full embodiment, including access to all 5+ of his human senses, until ALS began to weaken his physical abilities in his 20s and 30s. Would he have been able to make big discoveries in gravitational and theoretical physics without falling over? Or being able to move pen and paper? Or playing with ball models?

3. All major IQ tests for under 18s include physical object manipulation (like blocks, toys, chips, cards and other manipulatives for fine motor skills of the hands and fingers). For Wechsler this is the WPPSI and WISC. For Stanford-Binet this is the SB-5 and the older Form L-M.

Some further reading for interest:

Paper: The necessity of embodiment (2019, PDF).

LessWrong: Embodiment is Indispensable for AGI (Jun/2022).

More than a decade ago, the average human was a 28-year-old man from China: ‘He is Han Chinese so his ethnicity is Han. He is 28 years old. He is Christian. He speaks Mandarin. He does not have a car. He does not have a bank account.’ (NatGeo study cited by CBS, 2011).

The median human in 2024-2026 may meet these dot points:

A fair question in the age of millions of newly-minted AI experts! Alan is the ‘secret weapon’ for many AI labs, companies, and governments. He started his AI journey in the early 1990s, developing AI chatbots in QBASIC, followed by a degree in Computer Science with Psychology. He spent a decade leading applied human intelligence research, including as Chairman of Mensa’s gifted families. He rejoined the artificial intelligence fold with the launch of GPT-3 in 2020. He’s also known for things like:

– G7 (Group of Seven, global economic policy) referenced his GPT-5 report in their document building on The Hiroshima Framework.

– His visualizations were featured by professors at NeurIPS 2024.

– Apple leveraging his research in their new AI model paper for September 2024, using it as the foundation for their visualizations.

– Microsoft, RAND, and the European Commission, a few of the 10,000+ clients reading the regular advisory, The Memo, a Substack ‘bestseller’ in 147 countries.

– Ernst & Young, Ray White, USAA, and major government agencies consistently book him as their go-to speaker and educator for artificial intelligence topics.

– The largest asset managers in the world, with trillions of dollars in assets under management, use Alan for ongoing AI advisory.

– Fortune 500 companies import his AI expertise for in-person consulting engagements.

– Think tanks like Brookings Institution, BloombergNEF, RAND, and NBER frequently cite Alan’s research.

– NYU and other universities leverage visualizations and the popular Models Table, now detailing 500+ large language models.

– ‘What’s in my AI? (2022)‘ is said to be the most comprehensive analysis on GPT datasets, and was followed up by ‘What’s in GPT-5? (2024)‘ and ‘What’s in Grok? (2025)‘.

– More than 5 million early adopters watched Leta AI, across 67 episodes. Leta was powered by GPT-3 175B, a frontier model 2½ years before ChatGPT was released.

And much more: LifeArchitect.ai/about-alan

Milestones & justifications (most recent at top, skip to bottom ↓)

Google: ‘Gemini 4 Argon marks an important milestone accelerating the path toward physical AGI.’

Countdown note: Argon learns new physical skills by practicing in simulation, then self-corrects on real hardware through trial and reflection (58.9s ➜ 17.4s), a clear signal of adaptive learning from physical feedback. The final 1% now rests on a single demonstration of multi-step fine motor assembly, which looks imminent.

Announce,

cornhole video,

ERQA benchmark,

coding case study,

Models Table

Oct/2026 INFO: Unitree founder Wang Xingxing: GPT-7 next year may ‘completely break through the embodied AI problem’. Speaking at the opening of the 26th China International Industry Fair in Shanghai (12/Oct/2026), the Unitree Robotics founder said the ‘ChatGPT moment’ for embodied AI is accelerating, and could arrive in as little as one to two years. His threshold: when robots can complete 80% of tasks in 80% of unfamiliar scenarios using voice commands alone. This sharply revises his Aug/2026 (two months ago!) World Robot Conference timeline of ‘2 to 3 years at the fastest, 5 or 10 years at the slowest’. Wang credits GPT-6 for the shift: ‘We put it on all kinds of robots for all kinds of task tests, and the results are actually very good… The biggest mistake I made over the past decade or so was not believing in AI enough.’ He noted that factory pilots are underway at Unitree and other robot companies, with generalization and efficiency still needing improvement before large-scale deployment.

At the same event, Shanghai launched its new-generation intelligent manufacturing action plan, targeting 8,000 humanoid robots in factories and industrial robot density of 600 per 10,000 workers by 2028.

Countdown note: Wang’s 80/80 test (80% of tasks in 80% of unseen environments, voice-only instruction) is a useful industry yardstick that sits close to this page’s Woz test and IKEA milestones. Article (Chinese),

GPT-6 analysis,

Models Table

Oct/2026 INFO: Gemini 4 Argon hits lowest hallucination rate of any frontier model. On Artificial Analysis’s knowledge testing, Argon hallucinated on just 15% of the questions it couldn’t answer correctly, against 51% for GPT-6 Astra and 54% for GPT-6.1 Sol. When Argon doesn’t know, it says so 85% of the time. Accuracy is lower (50% vs Astra’s 63%), a trade made on purpose: the model now prefers ‘I don’t know’ to confabulation.

For comparison, the median human is far less careful with the truth. People tell an average of one to two lies per day, and lie in roughly one of every five social interactions lasting 10 minutes or more (DePaulo, 1996; Serota, 2010). That’s before counting honest confabulation: misremembering, overclaiming, and confidently filling gaps.

Countdown note: this meets (and likely exceeds) the median human standard, and clears the truthfulness milestone (grounded in an accepted version of truth without confabulation or hallucination). Eval,

DePaulo (1996),

Serota (2010),

Models Table

Oct/2026 INFO: Boston Dynamics new 13-DoF hands for Atlas. Four fingers (no pinky) and a 4-DoF thumb, up from 7 DoF. Slightly larger and stronger than a human hand, with tactile sensors across the fingertips and palm. Atlas can now pinch small objects, turn them in its hand, recover a slipping grip, and hold power tools while pressing the trigger. Boston Dynamics: ‘The ultimate general-purpose dexterous manipulator will not be an over-complicated hand, but one that is reliable and is capable of using tools.’

Countdown note: tool use and fitting parts together are the core motions of the IKEA furniture milestone. Paired with Gemini Robotics on Atlas, these hands may close the final gap.

Announce,

video

Sep/2026 INFO: Google DeepMind Gemini 4 Argon. New frontier model for long-horizon coding, enterprise knowledge work, and cyber defense, with the output limit expanded from 64K to 1M tokens. Google reports SOTA on real-world software engineering and #1 on the Vals Index (economic impact across finance, coding, legal, and tax). Internal results include beating a published quantum subroutine baseline by 40% in minutes, freeing 300+ TiB of datacentre memory via autonomous agents, and migrating C/C++ codebases (up to 800K+ lines for the Fuchsia Zircon kernel) to Rust.

Countdown note: Gemini 4 Argon is the likely foundation for the next generation of Gemini Robotics, developed alongside partners Boston Dynamics (Atlas) and Apptronik (Apollo). Pairing this model with those robotics stacks may be what closes the final gap in multi-step fine motor assembly. Announce,

Models Table

Sep/2026 INFO: Stanford + Caltech HomeBody: GPT-6 Astra drives a Unitree G1 humanoid directly, with no VLA in the loop. Instead of the usual three-part stack (System 2 VLM ➜ System 1 learned VLA ➜ System 0 controller), HomeBody lets a frontier VLM call a library of composable motor skills (navigate, pick, place, open drawer, pick from drawer) through structured tool calls. Deployment has three steps: the humanoid explores an unseen kitchen (iPhone video, D435i camera, LiDAR SLAM, joint poses, waypoints chosen by Astra), Astra then acts as a Real2Sim agent to build its own digital twin in Isaac Sim, and finally the robot takes everyday instructions. Demos: tidying a kitchen (gathering coffee bags onto the island, discarding spoiled milk and juice cartons across repeated trips), and ‘I forgot my medicine, can you get it for me?’, where the medicine is out of view, so HomeBody recalls stored keyframes to find the right drawer, opens it with the right hand, hands the medicine to the person, then discards a carton with the left hand. No environment-specific training data or additional policy learning. Bounded local retries (new grasp candidates, stance changes) handle failed grasps, and unrecoverable failures are reported back to the VLM to replan. Skills run on a single RTX 4090 laptop, with Astra called remotely. Limitations: Real2Sim setup time and API cost, reasoning latency pauses between skills, and finger-servo overheating during extended operation.

Countdown note: a frontier LLM exploring a strange room, building its own spatial memory, and orchestrating whole-body loco-manipulation zero-shot is another strong signal toward the Woz test and adaptive learning. The skills are coarse (grasp, place, pull), with no fastener or multi-part assembly, so the IKEA milestone remains open and the countdown holds at 99%.

Project page,

video,

repo,

Models Table

Sep/2026 INFO: GPT-6 Astra scores 80% in IKEA furniture assembly builds. Epoch AI purchased and assembled three IKEA items (STÄLL shoe rack, TONSTAD bed frame, GULLABERG 8-drawer dresser), photographing each build and deliberately introducing realistic mistakes, often continuing to build past the error to hide it. 60 photos total. Each model gets the assembly manual PDF, a zoom tool, and a Python interpreter, then must either list every step containing a mistake (with a correct description) or direct th