Bengio-linked researcher calls Astra AGI messaging mostly hype
He uses Astra often and finds it stronger than earlier OpenAI models but not a sea change toward AGI, so OpenAI's AGI messaging was mostly hype.
So looking back at the "AGI" messaging with which OpenAi announced the Astra (GPT 6) model, to what extent does it now sound like hype versus justified enthusiasm? It's a great model, I use it frequently (it feels like a peer to Fable 5.1 and Opus 5.5)... but while it does beat previous OpenAi models in significant ways especially for certain sorts of deep technical work, it does not feel like a sea change toward AGI So I would say: The AGi messaging about this model was mostly hype, really... I try to be open minded and was willing to give OpenAi some benefit-of-the-doubt because their stuff is super cool and useful and sometimes revolutionary, but just calling it as I see it... Looking back, basically the series of LLM model upgrades has been incremental in intelligence improvement rather than there being one huge leap, but if I had to look back at the biggest leaps so far it might be -- GPT4, which was the first model that basically made sense most of the time -- Claude Opus 3, which was the first LLM to be a genuinely decent writer. Also the first model smart enough in the right ways to power agentic loops like OpenClaw. -- DeepSeek-V3, the first model showing one could get near-frontier quality so much cheaper.. -- Opus 4.5 and ChatGPT 5.4 early this year ... the first coding models that clearly obsoleted most of my own efforts as a human coder Astra definitely makes fewer mistakes in complex math and has more polish and insight in complex software design than its OpenAi predecessors, but it doesn't seem as big a leap as some of these previous ones. As the AI ecosystem as a whole hurtles toward AGI, meaningful steps and contributions are coming from all over, and it's not wrong to say that each of these advantages is helping us build toward AGI. Astra among other things is helping my team's Omega agents more rapidly design and develop self-improving neural-symbolic agents trying to uplift themselves toward AGI/ASI -- and etc. etc. But even so, calling out Astra as the big uber special leap toward AGI seems in hindsight like very clearly mostly a bunch of hype... Not a shocking conclusion whatsoever, I just like to work through these things very clearly... There may before long come an AI release that genuinely is a sea-change leap toward AGI -- a bigger leap than the larger-leap-than-Astra LLM-model improvements I've listed above. I hope by then the AGI hype noise has not dulled the world to digesting and appreciating the genuine larger advances as they emerge... One can see the risk of a "boy who cried AGI" phenomenon here ;p