Kestrel Exams kestrel logo KESTREL EXAMSPRACTICE TO PASS

Kestrel Exams · AI

Curiosity and artificial general intelligence

Two claims today. One: the part of me that never grew up is a good thing. Two: the generation of AI OpenAI began rolling out on 3 September 2026 deserves to be called artificial general intelligence.

Big claim. You should expect me to explain it, and I should expect you to push back.

1 / 10

A simple question

"How many degrees of separation from Obama am I?"

Most likely three degrees of separation from Barack Obama, with a high probability of being no more than four.Google's AI, paraphrased. Encouraging.

That is an estimate. It is not a verified chain of people who could introduce me. I still have nobody to call.

What would it need to know to give me the chain?

2 / 10

Turning a vague question into a problem

Two dots, and the search between them

A route on a map. In real life the relationships have to be real, and the people willing to help. Standing next to someone in a photograph is not a phone number.

3 / 10

Where the argument begins

Pathfinding is an old problem. Look at everything around it.

Recognisewhat I actually wanted: a conversation, not a number.
Noticewhat information was missing, and choose a useful way to investigate.
Recoverwhen a promising lead is not supported, and try another approach.
Then tomorrowhelp with a completely different problem, without somebody building a separate program for each task.
4 / 10

Definitions

The standard I am asking you to consider

The practical standard (today's)Use knowledge across a broad range of tasks, learn what a new situation requires, and work toward a goal with limited step-by-step direction.
OpenAI Charter (stricter)Highly autonomous systems that outperform humans at most economically valuable work.
DeepMind, Levels of AGI (stricter)Measure breadth alongside performance, and treat autonomy as a separate scale.

Attempting many things, being good at them, and being trusted to do them alone are three different achievements. And none of this is a claim about feelings or an inner life. The argument is about what it can do.

5 / 10

What OpenAI reports

GPT-6 Astra, phased rollout from 3 September 2026

Works across computer applications Investigates questions Writes software Analyses data Produces finished documents
ARC-AGI-3, unfamiliar interactive environments
99.9%

A benchmark is evidence about performance on that test, not a certificate of competence at everything. What interests me is the combination: broad capability, tools for doing work, and the ability to adjust as the task develops. That is an interpretation we should test.

6 / 10

A better test than "did it sound smart"

Ask it to organise a campus event. Then check.

Before you say it did well

  • Does the schedule fit the room booking?
  • Does the budget add up?
  • Did it invent a speaker who never agreed to come?

Then change something

  • The room is unavailable.
  • The budget has been cut.
  • Can it revise the plan and keep every other requirement straight?

Try unfamiliar tasks. Check the work. Count the failures and the help required. Repeat in different areas. A serious claim has to survive more than a carefully chosen demonstration. Hold my argument to that standard too.

7 / 10

The child with the drawing

A willingness to begin before you know how to finish

What experience teaches us to say

  • "I don't know how to code."
  • "I wouldn't know where to start."
  • "Someone with more experience would have to do that."

What the same sentences become

  • What would I need to learn?
  • What could I build first?
  • How would I check whether it worked?

You still have the work to do. A polished answer you cannot evaluate is a reason to learn more about your subject. Having a good answer is not the same as having learned it. And a system can help me find a route to President Obama; it cannot make him want the conversation.

8 / 10

Your homework

One question you have been carrying around

1 · PickSomething you can check.
2 · AskAn AI for the first step, with enough context to be useful.
3 · InspectWhat it produces. Ask why it made that choice.
4 · ChangeA requirement, and watch what happens.
5 · NoticeWhat you learned, including where it needed you.

You might leave thinking I am too early to call this AGI. You might decide the term fits. Either way you will have more to contribute than a reaction to a headline, and you may have begun something you would otherwise have left sitting in your head.

9 / 10

Closing

I don't need you to tell me the drawing is good

Tell me what you see.

Ask me what I'm trying to make.

Show me where it doesn't work.

Enough enthusiasm to try something ambitious, and enough humility to keep learning when the first attempt falls short.

And if any of you actually know Barack Obama, please come find me afterward. I still want that chat.

10 / 10

The evidence

Is GPT-6 Astra artificial general intelligence? A graded answer

The presentation above makes the case in the speaker's own terms and asks you to test it. This article is the appendix: the framework, the published numbers, the demonstrations, the regressions, and a grade you can argue with.

Published 8 September 2026 · Kestrel Exams · Every claim below is linked to its source at the bottom of the page.

OpenAI released GPT-6 Astra on 3 September 2026 and described it as "the world's most intelligent and aligned model." At the launch event, Greg Brockman said that AGI has arrived. The written announcement is more careful: it talks about "a new generation of intelligence" and never uses the term. This page does what the announcement does not. It picks a published framework for artificial general intelligence, puts the evidence on the table, keeps the counter-evidence in view, and gives a grade.

What artificial general intelligence means

Artificial general intelligence (AGI) is the name for an AI system that can do a wide range of cognitive tasks at or above the level of a skilled human, rather than one narrow task like playing chess or transcribing speech. The trouble is that "wide range" and "skilled human" have never been pinned down by a single agreed test, so three working definitions get used in practice.

The OpenAI Charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work." It is a useful ambition and a poor yardstick, because nobody can enumerate most economically valuable work. ARC Prize, created by François Chollet, measures how efficiently a system acquires new skills on puzzles it has never seen; Chollet's stated position is that passing ARC is necessary for AGI but not sufficient. The Levels of AGI framework from Google DeepMind (Morris et al., 2023) is the one most researchers cite, because it turns the question from yes-or-no into a grid: how well does the system perform, measured as a percentile of skilled adults, and how general is that performance across kinds of task.

LevelPerformance barNarrow example (from the paper)General example (from the paper)
1 EmergingEqual to or somewhat better than an unskilled humanSimple rule-based systemsChatGPT, Bard, Llama 2 (2023)
2 CompetentAt least the 50th percentile of skilled adultsSmart speakers, WatsonNot yet achieved (2023)
3 ExpertAt least the 90th percentileSpelling checkers, image generatorsNot yet achieved
4 VirtuosoAt least the 99th percentileDeep Blue, AlphaGoNot yet achieved
5 SuperhumanOutperforms every humanAlphaFold, AlphaZeroArtificial superintelligence

The same paper adds a second scale for autonomy, from level 0 (no AI) through tool, consultant, collaborator and expert to level 5, AI as an agent. That scale matters later, because two of the features people are talking about most are about autonomy rather than raw capability.

What OpenAI published

The launch page reports benchmark results, and the following are the ones that bear on generality. ARC-AGI-3, the novel-reasoning benchmark: 99.9%, with human parity on 96% of levels. FrontierMath Tier 4, research-level mathematics: 98%, with the page crediting the model for improved bounds on prime gap theorems. GPQA Diamond, PhD-level science questions: 96.0%. ExploitBench, a cybersecurity exploitation benchmark: 100%, against 78.5% for GPT-5.6 Sol, and the model found two previously unknown zero-day vulnerabilities during testing. OSWorld 2.0, operating a real computer to finish tasks: 72.6%, in roughly 40 minutes per task versus 75 minutes for the previous model. OpenScore String Quartets, a symbolic music benchmark: 0.84.

The page also reports alignment numbers that matter for any AGI discussion: a 0% rate of exceeding authorized scope in agentic tasks, against 48% for GPT-5.6 Sol, and the model is three times less likely to misrepresent its own capabilities. The model is priced at $10 per million input tokens and $50 per million output tokens through the API, and is rolling out to ChatGPT Plus, Pro, Business and Enterprise.

OpenAI's launch video, from the official OpenAI YouTube channel.

The barrier, measured on our own work

Abstract benchmarks are easy to argue about. A barrier you hit yourself is not. In August 2026 our sister channel, A Wandering Mind, produced a 60-second science short and needed a music bed under the narration: a calm, plucked-string track that builds through the low end and leaves room for a voice. The brief was written carefully, with tempo, instrument character, arrangement and mix notes. The model of the day could turn that paragraph into audio. It could not reason about the structure of what it made.

Our production notes record what that cost. About eighteen candidate tracks were generated for one bed. Five deliberate revisions of the brief, each one pushing harder on the pluck rhythm, the ambient bed or the arc, all produced worse takes. Past a certain point, adding detail to the brief stopped helping and started hurting. There was no way to tell the model "the motif enters too late" or "the build should come from the bass, not the brightness" and have it fix that specific thing, so production became generate-and-pick, and the picking was done by a human ear. Our own measurements were no better: a beat-strength metric flagged the winning take as failing the brief's midpoint groove, and the person who had listened was right to overrule it.

That is the shape of the pre-Astra barrier. Not "the model cannot make music," but "the model cannot hold a musical structure in mind and work toward it."

What changed: Astra at the keyboard

The demonstrations that circulated after launch are different in kind. In the launch coverage, Astra plays a virtual keyboard live, choosing keys one at a time to produce a new piece rather than rendering a finished audio file. In independent testing it composed Bach-style chorales with zero voice-leading errors, which is the result that matters most. Voice leading is a rule system: four independent melodic lines that must move without parallel fifths or octaves, without collisions, and with each voice resolving where the harmony says it should. Satisfying those rules note by note is structural understanding, and it is what conservatory harmony courses spend a year teaching.

A working film composer reported being able to set up a three-minute loop in a DAW plugin and have Astra improvise over it for hours, adding one note per pass with chosen velocities, note lengths and pitches within a key, and described it as the first model that could do this. Another user prompted the model in LilyPond notation and had a playable virtual piano loaded with Brandenburg Concerto material in eleven minutes. None of this came from a music-specific product. The same weights that scored 98% on research mathematics and 100% on ExploitBench did it. That is the generality argument in one demonstration, and it is why the keyboard clip belongs in an AGI discussion rather than a music-tech roundup.

Planning and goals: from answering to finishing

Two features move Astra along the autonomy axis of the Levels of AGI framework. Goal mode in Codex became generally available on 21 May 2026 across the Codex app, IDE extension and command line. OpenAI's own description is that you "define an outcome and success criteria and let Codex keep working toward it." ChatGPT Work, announced on 9 July 2026, is described as "an agent that can take action across your apps and files, stay with a project for hours if needed, and turn a goal into finished work," breaking long projects into smaller steps and completing them independently, with scheduled and event-triggered tasks.

Those two features are the exact thing the music bed problem lacked. We could not hand a goal to the model and let it iterate against a success criterion; we could only prompt, listen and prompt again. On the DeepMind autonomy scale, that moves the system from level 3, collaborator, toward levels 4 and 5, expert and agent. The 0% authorized-scope violation figure on the launch page is what makes those levels usable rather than dangerous.

OpenAI's developer video, which covers Codex and the agentic features, is in the videos section below.

The videos

Every GPT-6 Astra video on OpenAI's own YouTube channel at the time of writing, plus the launch-coverage video the keyboard demonstration comes from. Play any of them here without leaving the presentation.

Introducing GPT-6 Astra: the most intelligent and aligned model in the world

OpenAI, 3 September 2026. The launch film: what the model is for, and the demonstrations OpenAI chose to lead with.

Introducing GPT-6 Astra for developers

OpenAI. Codex, Goal mode and the agentic features, which is where the "planning and goals" argument on slide 8 lives.

First impressions of GPT-6 Astra from developers

OpenAI. Early-access developers on what changed for them. Useful for the "from answering to finishing" point.

GPT 6 Astra is a freak (AI Search)

AI Search, launch coverage. This is the video the live keyboard demonstration on slide 7 is drawn from. Independent channel, not OpenAI.

What the evidence does not support

A grade is only worth something if it includes what went backwards. A hands-on review of Astra found that its writing quality dropped materially: on an editorial benchmark it scored 1,995 against 2,156 for GPT-5.6 Sol, a decline of roughly 80 Elo points, and evaluators described the prose as flat, lacking personality and easy to identify as machine-generated. One evaluator advised creative writers to avoid it. An AGI that is worse at the most common cognitive task in the economy than its predecessor is not Expert AGI, whatever it does in mathematics.

Three more cautions. OSWorld 72.6% is roughly human parity on computer use, which places that column at Competent, not Expert. ARC-AGI's creator maintains that a high score is necessary for AGI and not sufficient, so 99.9% closes one door rather than opening all of them. And the AGI claim itself came from a stage; the launch page, the document OpenAI chose its words for, does not make it.

The verdict

DomainEvidenceLevel of AGI
Research mathematicsFrontierMath Tier 4 98%; new bounds on open problems4 Virtuoso to 5 Superhuman
Novel abstract reasoningARC-AGI-3 99.9%, human parity on 96% of levels3 Expert to 4 Virtuoso
CybersecurityExploitBench 100%; two zero-days found4 Virtuoso
Symbolic music compositionChorales with zero voice-leading errors; live keyboard; DAW improvisation3 Expert
Science knowledgeGPQA Diamond 96.0%3 Expert
Operating a computerOSWorld 2.0 72.6%, about human parity2 Competent
WritingAbout 80 Elo below GPT-5.6 Sol on an editorial benchmark2 Competent, regressed
AutonomyGoal mode; ChatGPT Work; 0% scope violationsAutonomy level 4, approaching 5

Grade: Competent AGI, with Expert-to-Virtuoso spikes. The published evidence supports level 2 across a wide range of non-physical tasks, and levels 3 and 4 in mathematics, abstract reasoning, cybersecurity and symbolic music. It does not yet support Expert AGI across the board, because writing regressed. When the Levels of AGI paper was written in 2023, no general system had reached Competent. That rung is now occupied, and the gap between it and the next one is smaller than the gap it just crossed.

How this page was built, and what we did not verify ourselves

The benchmark figures are taken from OpenAI's launch page. The writing regression, the Bach chorale result and the "AGI has arrived" quotation come from a published hands-on review and are reported as that review's findings. The DAW improvisation account comes from a working composer's published write-up. The LilyPond virtual piano demonstration comes from press coverage of user demos. The live keyboard demonstration is from a launch-coverage video linked below; we did not run that demonstration ourselves. The music bed figures are from our own production notes for an August 2026 short. Where a claim rests on one source, the table above says so by naming the evidence rather than the conclusion.

Questions people ask

Is GPT-6 Astra AGI?

On DeepMind's Levels of AGI ladder, yes at the Competent level, with Expert-to-Virtuoso results in several domains, and no at the Expert level across the board. OpenAI's written announcement does not claim AGI; its president did, on stage.

Why does a piano demo count as evidence about intelligence?

Because it shows rule-governed structure being satisfied note by note in a model that was not built for music. Audio generators existed for years before Astra; what they could not do was hold a structure in mind and work toward it, which is exactly what our own production notes show us failing to get out of them.

What is the difference between Goal mode and ChatGPT Work?

Goal mode lives in Codex, the coding agent, and takes an outcome plus success criteria. ChatGPT Work is the general agent in ChatGPT that takes a goal, plans the steps, and works across your apps and files for hours. Both are autonomy features; neither is a new model.

Will the grade change?

Yes, and this page will be updated when it does. The two things that would move Astra to Expert AGI are an independent writing evaluation at or above the 90th percentile and Expert-level results on computer use. Both are measurable.

Sources

  1. OpenAI, "GPT-6 Astra: A new generation of intelligence" (launch page, 3 September 2026): openai.com/index/gpt-6-astra
  2. OpenAI, "Safety overview: GPT-6 Astra" (3 September 2026; confirms the announcement date and the Critical cybersecurity capability designation): openai.com/index/safety-overview-gpt-6-astra
  3. OpenAI, "Introducing GPT-6 Astra: the most intelligent and aligned model in the world" (video): youtube.com/watch?v=1QNsdr-Qx_I
  4. OpenAI, "Introducing GPT-6 Astra for developers" (video): youtube.com/watch?v=bOC3DisEOfg
  5. OpenAI, "First impressions of GPT-6 Astra from developers" (video): youtube.com/watch?v=-TTyyY3VWh8
  6. Morris, Sohl-Dickstein, Fiedel, Warkentin, Dafoe, Faust, Farabet and Legg, "Levels of AGI for Operationalizing Progress on the Path to AGI," Google DeepMind, 2023: arxiv.org/abs/2311.02462
  7. OpenAI Charter: openai.com/charter
  8. ARC Prize, "What is ARC-AGI?": arcprize.org/arc-agi
  9. TechFlow, "GPT-6 Astra Hands-On Review" (writing regression, Bach chorale result, Brockman quotation): techflowpost.com/en-US/article/33795
  10. Sound for Movement, "GPT-6 Astra for Professional Musicians" (DAW improvisation): soundformovement.com
  11. The News International, "OpenAI's GPT-6 Astra shows off new skills in music, games and complex tasks after launch" (LilyPond virtual piano): thenews.com.pk
  12. AI Search, "GPT 6 Astra is a freak" (launch coverage including the live keyboard demonstration): youtu.be/Ji4amrxrzVM
  13. OpenAI Help Center, ChatGPT release notes, 21 May 2026 entry on Goal mode: help.openai.com
  14. OpenAI, "ChatGPT is now a partner for your most ambitious work" (ChatGPT Work, 9 July 2026): openai.com/index/chatgpt-for-your-most-ambitious-work
  15. Kestrel Ventures production notes for the A Wandering Mind photosynthesis short, August 2026 (internal; music bed candidate counts and brief revisions).

Studying computer science, and want the practice to match?

Kestrel Exams has free, offline practice tests for Java, Python, Discrete Math and Calculus, written in course order. No account, no upsell.