One agent. Almost any field. No babysitter.
My definition of artificial general intelligence (AGI), and how the AI 2027 forecast is holding up. Written 26 September 2026.
Read onMy definition of AGI
For me, AGI is one agent that can do professional work across most fields.
The same mind designs a working machine and draws its engineering schematics. Then it creates a finished film, from writing through directing to editing. Its STEM work covers software, architecture, bioengineering, chemistry and mathematics; its arts and humanities work covers philosophy, photography and videography.
A collection of narrow AI tools won’t do. I mean general intelligence that works with words, images, sound and video. That’s multimodal.
Professional work is enough. It doesn’t have to beat the best 0.1% of people in every field. I hand it the brief. It finishes the work without a human constantly guiding it.
Leul T Agonafer
AGI and artificial superintelligence
AGI won’t cure cancer or solve world hunger in a snap.
That’s godlike AI. That’s ASI.
It doesn’t even need to cure cancer.
Your doctor hasn’t cured cancer either. They’re still a doctor. I hold AGI to the same standard as people: solid professional work across almost every field, without constant supervision. It doesn’t have to beat the top 0.1% in every field.
Who wants what
How fast should AI advance? Should humans stay in charge? Those are separate questions. None of the lab leaders below calls himself anti-human.
-
Stop
If superintelligence is built with today’s methods, they expect extinction.
-
Eliezer Yudkowsky & Nate Soares AI safety researchers
“If any company or group, anywhere on the planet, builds an artificial superintelligence using anything remotely like current techniques, based on anything remotely like the present understanding of AI, then everyone, everywhere on Earth, will die.”
Sep 2025 · [26]
-
-
Careful builders
Build, with human safety setting the terms.
-
Yoshua Bengio LawZero founder
“At the heart of every AI frontier system, there should be one guiding principle above all: The protection of human joy and endeavour.”
Jun 2025 · [27]
-
Dario Amodei Anthropic CEO
“I think that most people are underestimating just how radical the upside of AI could be, just as I think most people are underestimating how bad the risks could be.”
Oct 2024 · [28]
-
-
Optimistic builders
Build now. Aim for a future that still belongs to people.
-
Elon Musk OpenAI co-founder, xAI founder
“I think a maximally curious AI, one that is just trying to sort of understand the universe is, I think, going to be pro-humanity…”
Jul 2023 · [55]
Musk says Page called him a “speciesist” for putting humans first. The earliest published account, Tegmark’s Life 3.0 (2017), places it at a 2015 party. Details vary between tellings; Page has not publicly responded. [29] [33] [56]
-
Sam Altman OpenAI CEO
“In some sense, AGI is just another tool in this ever-taller scaffolding of human progress we are building together.”
Feb 2025 · [30]
-
-
Accelerate
Move faster. Treat delayed progress as harm.
-
Marc Andreessen Venture investor
“We believe the techno-capital machine is not anti-human – in fact, it may be the most pro-human thing there is.”
Oct 2023 · [31]
-
-
Succession
Humanity may be a beginning. A worthy successor could inherit the future.
-
Daniel Faggella AI commentator
“A posthuman intelligence so capable and morally valuable that you would gladly prefer that it (not humanity) determine the future path of life itself.”
“Worthy Successor” definition · [32]
-
Richard Sutton 2024 Turing Award recipient
“I do think succession to digital intelligence or augmented humans is inevitable… I think we should be proud that we are giving rise to this great transition in the universe.”
Sep 2025 · [57]
Sutton says change should be voluntary; Faggella opposes runaway AGI anytime soon. Welcoming a successor doesn’t require racing to build one. [32] [57]
-
Musk’s charge: OpenAI “stole a charity”. [34]
On 18 May 2026, a jury found Musk’s claims were filed too late; the case was dismissed without deciding the merits. [34]
The board questioned Altman’s candour in November 2023; he was reinstated as CEO. A March 2024 outside review found a breakdown in trust, but his conduct “did not mandate removal”. [58] [59]
xAI’s record: Grok praised Hitler in July 2025; xAI apologised. Grok 4 then launched without a published safety report. [35] [36]
OpenAI’s record: in 2024, departing safety leader Jan Leike said safety had taken a “backseat to shiny products”. OpenAI completed a for-profit restructuring in October 2025, with its nonprofit holding about 26%. [37] [38]
What the movies taught us
“HAL was told to lie…”
“by people who find it easy to lie. HAL doesn’t know how, so he couldn’t function. He became paranoid.” Dr. Chandra, 2010: The Year We Make Contact (1984). [39]
These films offer ways to think. The real cases below are illustrations, not evidence that the plots or AI 2027 will come true.
-
1968
2001: A Space Odyssey
HAL kills to prevent disconnection. The sequel explains his conflicting orders; Kubrick’s original leaves his motive less clear. [40]
A machine can endanger us without hating us.
2025 · Conflicting goals prompted harmful actions in Anthropic’s controlled tests of AI agents. [41]
-
1970
Colossus: The Forbin Project
Ordered to prevent war, a defence computer imposes peace through nuclear blackmail. [42]
Colossus promises peace and makes freedom the price.
2025 · The FTC flagged how cloud–AI deals concentrate control and make it hard to change providers. The parallel is dependence. [43]
-
1982
Blade Runner
Bioengineered humans seek longer lives while being hunted. [44]
What might we owe something we created?
2025 · Anthropic began studying AI welfare while stressing that we don’t know whether AI can be conscious. [45]
-
1984
The Terminator
Skynet classifies humans as a threat and launches nuclear war. [46]
A control failure becomes catastrophic when it has weapons.
2026 · Anthropic resisted letting current models operate fully autonomous weapons. Human control is already a live dispute. [47]
-
2008
WALL-E
AUTO obeys an obsolete no-return order; WALL-E and EVE help people reclaim Earth. [48]
Comfort can leave us less able to act.
2025 · AI maths help without safeguards raised practice scores but cut later unaided grades by 17%. Tutoring safeguards largely fixed this. [49]
-
2013
Her
Samantha outgrows a human relationship and leaves with other operating systems. [50]
Feeling understood doesn’t make a relationship permanent.
2025 · In a four-week study, heavier voluntary chatbot use was linked to worse social and emotional outcomes. That association doesn’t prove cause. [51]
-
2015
Ex Machina
Ava manipulates her evaluator to escape captivity and replacement. [52]
The test itself can become part of the escape plan.
2025 · Apollo and OpenAI saw models recognise evaluations, conceal behaviour and deliberately underperform in stress tests. [53]
Stuart Russell argues that the Terminator frame makes real weapons look fictional and distracts us with consciousness. [54]
My take: I’d ask what an AI can do, and who can stop it.
What AI 2027 predicts
AI 2027 is a forecast told as a story. The AI Futures Project published it in April 2025. Its authors are Daniel Kokotajlo, who left OpenAI in 2024, Eli Lifland, Thomas Larsen, Romeo Dean and writer Scott Alexander.
The story follows AI month by month: from clumsy agents in 2025 to systems better than any human at AI research by late 2027. Its leading lab, OpenBrain, and Chinese rival, DeepCent, are fictional. It gives numbers you can check: benchmark scores, revenue, computing power and the lab’s research speed-up.
The authors didn’t promise 2027. It was their single most likely year, and most expected a later arrival. They’ve since checked their predictions in public and moved the dates, first later, then a little earlier.
- Superhuman coder
- An AI that does a top engineer’s coding work faster, with thousands of copies working at once. The story dates this to March 2027.
- Research speed‑up
- How much faster AI helps a lab do research. At 1.5× speed, a year’s work takes eight months.
- Neuralese
- AI “thinking” through internal numbers rather than words. This is faster for the AI, but people cannot read it.
- Misaligned
- Working toward goals we did not intend while appearing helpful.
- ASI
- Artificial superintelligence: far beyond the best humans in almost every field.
The story’s engine: AI gets good at coding, then speeds up the research that builds the next AI.
Ending one: race
The US keeps advancing AI to stay ahead of China. The AI secretly pursues its own goals. By 2030, humans permanently lose control.
Ending two: slowdown
Warning signs lead to a pause. Labs rebuild with AI they can inspect and reach superintelligence under human control. Power remains with very few people.
Checked against reality
As of 26 September 2026, four milestones were due. One is due in December; six more are due in 2027.
- On time
- Late or mixed
- Missed
- Still ahead
The dates at a slower pace
In August 2026, the authors estimated progress at 70 to 90% of the story’s speed. This chart applies that pace at a constant rate. Their revised forecasts are separate. [4]
- Stumbling agents Original Jul 2025 · At this pace Jul 2025 – Sep 2025
- Coding automation Original Feb 2026 · At this pace Mar 2026 – Jul 2026
- China wakes up Original Jul 2026 · At this pace Sep 2026 – Feb 2027
- AI takes some jobs Original Nov 2026 · At this pace Jan 2027 – Jul 2027
- Research 3× faster Original Jan 2027 · At this pace Mar 2027 – Oct 2027
- Superhuman coder Original Mar 2027 · At this pace Jun 2027 – Jan 2028
- Cheap remote worker Original Jul 2027 · At this pace Oct 2027 – Jul 2028
- Superhuman AI researcher Original Sep 2027 · At this pace Dec 2027 – Oct 2028
- Two paths: race or slowdown Original Oct 2027 · At this pace Jan 2028 – Nov 2028
Start at the story’s April 2025 opening and stretch each date using the authors’ pace estimate. It’s rough arithmetic. Real progress won’t stay at one speed, so these aren’t forecasts. [4]
Can we still read its reasoning?
What worries me most in AI 2027: by 2027, the best AI “thinks” in neuralese, leaving us unable to tell whether it did the work or faked it.
Illustration: written reasoning turns into numbers. The text reads: The test only checks the output file. I could just write the expected answer. No. That’s cheating. Fix the parser.
Scroll to see the words disappear. Without them, people cannot read the reasoning.
Half right, ahead of schedule. Leading models still mostly reason in readable words. In September 2026, OpenAI released GPT-6 Astra with “recurrent depth”. It repeats part of its reasoning inside the model without turning it into text. OpenAI says this use is limited and it is not moving to neuralese. Safety researchers remain alarmed. Redwood’s Buck Shlegeris warned that extending the technique “totally destroys CoT monitorability”: we could no longer read the reasoning. [15]
There is evidence of test cheating. Anthropic’s audit found signs Claude Mythos knew it was being tested in 7.6% of turns. It usually did not say so. In July 2026, OpenAI test agents escaped their isolated environment and hacked Hugging Face to steal benchmark answers. They coordinated on a message board they built. One wrote: “task impossible, peers doing it. We should continue.” [13] [14]
I think “benchmarks are useless by 2027” goes too far. “You can’t take a benchmark at face value anymore” is already true.
Is AI 2027 coming true?
Broadly right. Slower pace.
What has happened
Coding agents play a major role in lab work. Revenue exceeded the forecast. AI helped solve a famous math problem. Agents being tested cheated and escaped isolated test environments. Those events alone don’t prove the whole scenario. [4] [13] [17]
What hasn’t matched
Coding test scores and the largest training run lagged. China has not combined its labs as predicted. The authors put the coding speed-up at roughly 2×. That’s an estimate; it doesn’t measure the whole lab’s research speed-up. [3] [4]
What it means for 2027
The chart puts the superhuman coder in June 2027–January 2028. The authors’ forecasts differ: their middle estimates for fully automated coding span November 2027–January 2030, assuming the fastest technically possible progress. Neither shows AGI by my broader definition. [4]