N8Programs 讨论 Astra 算不算 AGI

N8 Programs @N8Programs原推发布于 2026.09.07

在 X 上打开原推

Some thoughts on whether Astra is AGI. AGI is hard to define - but I particularly like @fchollet's definition - that we'll have AGI when "the exercise of creating tasks that are easy for regular humans but hard for AI becomes simply impossible." By this standard, across the board, Astra is not AGI - it (likely) cannot perform long-horizon embodied tasks (making a cup of coffee, assembling a piece of IKEA furniture), could not play real-time multiplayer games like CS:GO or Overwatch, and lacks audio input (which makes it obviously subhuman at a wide variety of audio tasks - even stuff as simple as transcribing words from audio w/out whisper)! However, I think this is slightly unfair to Astra - it exists in a world of much more limited I/O than humans do - it takes text and images as input, and can output text - some of which is tool-calls. And in this regime, I think a case can be made for Astra being AGI-like - i.e. its very hard to find a task that can be done on a computer, the inputs to which are text and images, and the desired outputs of which are text and images, where Astra is *worse* than the average human who is familiar with computer use. A prominent example of this is of course ARC-AGI-3, which Astra essentially saturates (fewer actions than the median human on 96% of levels). It crushes the human baselines on ARC-AGI-1 (98.5% vs 64.2%) and 2 (95.0% vs 60%). Astra is also superhuman (as in better than a median human in America/the world) on any form of STEM/classicailly verifiable knowledge task one can think of - the benchmarks its evaluated on typically show it massing or surpassing human *expert* performance. I'd be highly surprised if one could find a broad, non-gotcha task that showed otherwise. Some evidence that leads me to believe this below: GPQA - Astra gets 96%. PHD Expert gets 69.7% (81.3% among original validators). BrowseComp - Astra gets 91.5%. 29.2% was the average human baseline. HealthBench Professional - Astra gets 63.4 points, versus 43.7 points for human experts with internet access and no time limit. Most other benchmarks share this pattern, afaik. This isn't to say for a domain-specific task STEM/verifiable knowledge task, there exists no human that surpasses Astra - I'm sure there are many - but the average human surpassing Astra? Likely not. Furthermore, Astra appears to be a generational leap in computer use - it can get diamonds in Minecraft (https://x.com/wuyang_zhou/status/2096548032128942370), and beat Portal (https://x.com/cozyblazex/status/2096383114851533097). For the latter, the game pauses while Astra thinks - so its not entirely human-matched - but this is inevitable, given the way Astra is bound to the world is through a strict inspect image -> call tool -> inspect image loop (one that it is trained to use quite well!), instead of the more continuous interaction humans have between sensory input, motor control, and higher-level thought. Similarly, it can also use Blender (https://x.com/aollivier82/status/2096226819401801896 for a good example) with far better acuity than the average person (though it obviously pales in comparison to actual 3D artists). I'd say that Astra is likely better than the average person at computer-use tasks. Astra can also do basic robot control tasks - like moving an object into a cup (https://x.com/chooi_jeq/status/2096064315115839904), manipulating a block (https://x.com/k7agar/status/2096590458369810679), and (partially) clean a surface with a cloth (https://x.com/k7agar/status/2096593654320341027). Again, Astra has no native output channel for robotics beyond its standard image -> tool call -> image abilities. I'd suspect what Astra achieves here is similar-ish to what an average human would be able to do if they were hooked up to robot arms they could only control via GUI. This is what I'd expect to see from an AGI-like system produced with todays methods - a shit ton of crystal knowledge + reasoning ability that allows it to outperform even human experts on domains that it was trained heavily on (STEM tasks, some knowledge work, etc.), plus human-like general reasoning abilities in more novel domains (robotics control, general CUA, and ofc ARC-AGI). This is sufficient for me to think Astra "AGI-like" under @fchollet's definition if you restrict the domain of tasks to those that have text/image inputs and text/image outputs, and can be done via computer - and if average human means 'random average Joe' and not 'average skilled expert.' However, I'd hesitate to call Astra complete AGI in the sense of one-to-one human replacement, even under the aforementioned domain restriction. Mainly because there remain a lot of open questions about Astra's ability over very long time horizons. Could OpenAI keep making research at the progress it does today if it replaced all employees with Astra instances? Likely not. Similarly, there exist edge cases where Astra doesn't learn in the exact same way humans do - starting strong, but not learning as much over time (https://x.com/EpochAIResearch/status/2095602826357670001 - though this is a single human run ). This plays into the larger question of whether compaction or the more novel note-taking system Codex now uses is enough for continual learning at the human-level. Finally, it's unclear whether Astra could just automate a remote white collar job - as far as I know, no one has tried, and it hasn't been out long enough to test (I'd guess maybe, at slightly below median worker-on-the-market level, with heavy jaggedness). In summary: Astra has a lot of AGI-like qualities - in the sense that it outperforms the average person on almost anything in the realm of text/images -> text/images or tasks that require computer use (note that this is obviously impossible to prove, but existing results should update your prior toward it). It's not a general AGI in the sense it can do *any task a human can*, as it lacks the necessary senses and output modalities to do so. However, it appears to exhibit human-like levels of fluid intelligence in the domains for which it is designed (judging by its CUA abilities and ARC-AGI results) - meaning it may be able to perform at a human level in the aforementioned missing modalities if given the suitable tools for input/output. However, it remains unlikely that Astra matches humans on extremely long-horizon tasks, and exhibits the collective dynamics that humans do in terms of research ability (hence why OpenAI is not just a bunch of Astra instances). In summary of the summary: Astra is very AGI-like for any CUA task that has a clearly defined scope, and can apply human-like fluid intelligence to these tasks. It's unclear, and unlikely, that Astra generalizes this to extremely long time horizons in the way that would be necessary to fully substitute for human labor. In summary^3: AGI-like thing achieved with normal caveats.

这条是引用着另一条写的：x.com/epochairesearch/status/2095602826357670001