Not sure if this is a hot take or not but…
If senior people central to Gemini’s current strategy are leaving, and that strategy has produced a frontier model now delayed to July, then their departure potentially opens it up for a different direction rather than simply weakening it.
And… arguably that change is already evident. For example 3.5 Pro delay, compute for one of Shazeer’s projects reassigned to a GDM London team, Adler and Pritzel (both core Gemini contributors) heading to Anthropic. Hints, to me at least, that this is (potentially) a symptom of a wider change rather than its cause.
@OfficialLoganK Is comparing (the soon to be released🤞) Gemini 3.5 to Fable 5 the wrong way of framing things, ie are these models optimised for different things entirely?
Really interesting video, thank you. On the 3GB part... I may be misunderstanding, but if the genome stores a compressed learning rule rather than weights, could it just be that evolution pretrained our optimiser? In which case the "not enough space" point seems to strengthen the evolution objection rather than refute it.
@demishassabis Demis was a chess prodigy. Though it's never easy to lose top talent, he knows you sometimes sacrifice valuable pieces on the way to a bigger strategic win.
Round 2. Tried with a stricter prompt to reproduce it faithfully, no homage/variation.
Gemini 3.1 Pro: just reproduced the composition without pushback. I can imagine this titled as: 'Man walks through hurricane'.
GPT-5.5: insisted it couldn't do "exact" and produced something akin to a cyclops bat with doors for wings facing two pawns.
Claude Opus 4.8: naturally refused, but does its best Droste effect, so more the mechanism than the scene.
Personally I sense future LLMs may produce some really interesting Escheresque SVG 'art'
A more cultured SVG benchmark than a BMW. The frontier-model final boss: M.C. Escher's Print Gallery! If you've not seen the original, check that out first to give the LLMs some slack!
Round 1. I asked: "make an SVG of Escher's Print Gallery." Alas copyright compliance turned the request into a "homage" rather than a replica.
GPT-5.5 (high reasoning): the recursive gallery with the labelled "void". Feels the closest.
Claude Opus 4.8 (Max): the domed ceiling redolent of the Pantheon in Rome.
Gemini 3.1 Pro: minimalist!
@dioscuri If you imagine looking back 15-20 years from now, are there tv shows today that you can imagine will evoke a similar feeling in the future. Or to put another way, are there programs you enjoy today that you think will age badly. Is it possible to even know!
I can imagine GPT 5.6 release being delayed as a result of current US govt restrictions, i also wonder does this current situation also delay Gemini 3.5 release that Google previously slated for this month.
@emollick This is very reflective of many large orgs at present. Though I think i fork on the where it’s going - my expectation is the speed of transition for large orgs will likely not keep up with new start-ups with thousands of goal focused agents. Gradual, then all at once.
@DanielleFong@alexolegimas Although I think they are ‘real’ the video looks AI generated, so wouldn’t be confident of the veracity of the actual video. How effective they would be in terms of battery duration, control range etc limits the immediate practicality.
@bcherny 1st, thank you for Anthropic's amazing work! I had some ideas for the Claude app:
custom filters (eg a toggle for Deep Research).
scratchpad within chat (like /btw in CC) for tangent questions, footnoting them at the bottom so they don't pollute the main context window.