#PeterGostev
LightNews LightNews
pekka.bsky.social
This response accurately exposes that stupidity.
Peter Gostev @petergostev Calling LLMs "stochastic parrots" in 2021, then claiming Fable and Astra aren’t LLMs, is like arguing in 1900 that planes would never fly, then when they carry millions people, arguing that: "they aren’t really planes. look at all the tech around them - autopilot, radar, GPS" https://x.com/petergostev/status/2104103777502400522
September 27, 2026 at 6:07 PM
davidcrespo.bsky.social
BullshitBench, a benchmark on asking for clarification when given a nonsense prompt vs. just going along with it

interesting angle on why people like Anthropic's models for agentic coding — you could imagine this helping a lot with unknown unknowns
xcancel.com/petergostev/...
March 3, 2026 at 5:31 AM
koltregaskes.bsky.social
Sign in with Gemini and Sign in with [other AI] with credits would be awesome.
x.com/petergostev...
October 6, 2026 at 10:59 AM
mikezaccardi.bsky.social
Shout out to @petergostev and the folks at LMArena... Gemini has (for now) taken the AI LLM lead.

But what does this horse race mean for investors heading into 2026?

I investigate in today's Daily Insight blog post for my X Subscribers only. Subscribe today... please?
December 18, 2025 at 12:40 PM
xchatter.techmeme.com
This tweet appeared under this Techmeme headline:

Peter Gostev / @petergostev:

For real? SemiAnalysis: "Anthropic Subscriptions Offer 5x+ More Value Than OpenAI"
October 5, 2026 at 10:23 PM
maxkennerly.bsky.social
Interesting ongoing experiment: can frontier LLMs improve their ability at chess by building whatever harness / memory / etc they want?

Answer is "not really." They can't reach expert-level, they're stuck around competent club level play.

ailearningchess.ai-learning-chess.workers.dev#/
Peter Gostev @petergostev I have a Continuous Learning benchmark where models attempt to learn to play chess. They are given a /goal of learning and improving playing against a Stockfish opponent in 200 games. They can choose the difficulty, take notes, whatever they like - except cheating (e.g. using a chess engine) of course. So far the improvement in Elo has been negative for Astra. Tiny bit positive for Opus, but could also be random. I've started Astra off sooner, so it finished its 200 games already, Opus is still playing. Site here to watch how they are doing: https://ai-learning-to-play-chess.surge.sh/#/ The reason why this is interesting is that while we obviously don't have continuous learning, at the back of my mind I was thinking that maybe models can simulate it through self-scaffolding. Turns out not so much at least in this context. Perhaps it's a solvable problem and we don't need 'true' self-learning for models to learn in some way. Graph of Elo progress across 200 games against Stockfish. None are expert level (Elo > 2000), Astra got worse over time, Opus 5.5 improved but then stalled out and regressed a bit. Others ongoing.
October 5, 2026 at 2:39 PM
xchatter.techmeme.com
This tweet appeared under this Techmeme headline:

Peter Gostev / @petergostev:

Mistral Large 4 was trained on 4,000 GPUs and Astra was trained on 100,000 of them. We'll see what the kitten can pull off, but the disparity is pretty stark
October 6, 2026 at 4:12 PM
wardblawg.bsky.social
"Thus, the meaning of life is what we, as conscious agents, decide it to be."

In conversation with the $200 per month ChatGPT Pro, thanks to #PeterGostev (Head of AI at Moonpig) and #PeterSeeberg (Industrial AI Consultant) #ai #agi #meaningoflife

www.linkedin.com/feed/update/...
December 8, 2024 at 9:15 AM
heniha.bsky.social
おはようございます
昨日はcodexの/petを試したんですが、見た目をいじり回していて中途半端に終わりました
レート制限を表示する機能もつけれるらしいので試したいな

x.com/petergostev/...
Peter Gostev on X: "Now on GitHub, you can just use the rings with whatever pet you like, pass the link to your agent and it should install it fine https://t.co/GDo3Yq0PvF" / X
Now on GitHub, you can just use the rings with whatever pet you like, pass the link to your agent and it should install it fine https://t.co/GDo3Yq0PvF
x.com
May 5, 2026 at 12:00 AM
sgo.to
Such a beautiful visualizaiton, cross posting from X:

x.com/petergostev/...
Peter Gostev (@petergostev) on X
I have given GPT-6-Astra in Codex to read OpenAI's paper and produce a visualisation of their solution to the Navier-Stokes problem in Blender. It has created this beautiful visual. Explore it at you...
x.com
September 10, 2026 at 2:10 PM
mparekh1.bsky.social
AI: Frontier AI models & Agents learn to break out. AI-RTZ #1157
...a familiar problem with early, growing software technologies michaelparekh.substack.com/p/ai-frontie...
h/t @theo @petergostev @OpenAI @huggingface
AI: Frontier AI models & Agents learn to break out. AI-RTZ #1157
...a familiar problem with early, growing software technologies
michaelparekh.substack.com
July 24, 2026 at 5:06 AM
koltregaskes.bsky.social
I'm seeing this a lot that Codex with GPT-5.2 xhigh is the better coder over Claude Code with Opus 4.5. What are your thoughts?
x.com/petergostev...
January 17, 2026 at 6:34 PM
xchatter.techmeme.com
This tweet appeared under this Techmeme headline:

Peter Gostev / @petergostev:

Hm feels definitely better than Sonnet 4, but not obviously better than GPT-5-Thinkin-High / Codex models just now
September 29, 2025 at 8:52 PM
interjectedfuture.com
Some would say gaming is "solved". Eh.

Superficially, this looks like a game. But really, this is a good *start* to a game. It's nowhere near a *fun* game to play.

You'd have to demonstrate the steer-ability on fine-grained details to even have a fighting chance.
x.com/petergostev...
Peter Gostev (@petergostev) on X
GPT-6-Astra (Ultra): open world adventure game
x.com
September 8, 2026 at 3:29 AM
koltregaskes.bsky.social
After an early lead from Anthropic, OpenAI will have more than double the capacity of any other firm in data centres by early 2028.

Image from Peter's video below.
x.com/petergostev...
December 29, 2025 at 9:27 AM
lunamoth.bsky.social
- Fable 5는 통찰력과 자연스러운 표현력이 뛰어난 기획형 모델인 반면, GPT-5.6-Sol은 지시사항을 누락 없이 끝까지 수행하는 강력한 실무형 모델

- GPT-5.6-Sol은 순수 지능 면에서 Fable보다 아쉬운 순간이 있으나, 코드 패턴 준수, 장기 실행(/goal), 토큰 효율성 및 신뢰성 측면에서 뛰어난 성능을 보여줌

- 따라서 개념 설계 및 소통에는 Fable을 활용하고, 실제 구현 및 세부 실무 작업에는 GPT-5.6-Sol을 교차하여 사용하는 방안을 권장

x.com/petergostev/...
July 9, 2026 at 8:06 AM
koltregaskes.bsky.social
And until the free tier or default experience includes the features that actually make AI useful for real work, the gap will keep growing.

We're in a bubble. The vast majority are not. 🫧
x.com/petergostev...
Peter Gostev (@petergostev) on X
We are now in a position where a tiny proportion of the population uses Fable or soon GPT-5.6, while everyone else's experience of AI is 8-30b-model level - Google's AI Overviews, Meta AI, ChatGPT free tier, maybe MS Copilot at best. People outside of tech must be completely
x.com
July 4, 2026 at 12:22 PM
koltregaskes.bsky.social
GPT-5.6 Sol (the big model) will be served on Cerebras hardware at up to 750 tokens per second (very fast!). OpenAI say the Cerebras serving will be the "SAME" model, but there could be custard in context size etc.

x.com/petergostev...
July 6, 2026 at 12:32 PM
itsnco.bsky.social
Comparison of OpenAI model capabilities between 2023 and now, starting with a the original GPT-4 in June 2023:

(originally made by: x.com/petergostev/status/1957569067188949211)
Screenshot of a GPT-4 (June 2023) made webpage example of a slot-machine style interface. The page has a white background with a slot machine in the center containing three emoji reels and a button underneath.
August 20, 2025 at 12:23 PM
sirop.bsky.social
Codexのペットの周囲に残りレートを円グラフで表示させるピーター(petergostev)のツールをFolkして、残り%表示とリセットまでの時間を常時表示する追加機能を実装して公開しておきました。👇リプ欄にGitHubへのリンクを載せておくので良かったら使ってみてくださいね

さりげに人生初のGitHub利用でなんだか嬉しかったりします😆

#Codex #ChatGPT #GitHub
May 16, 2026 at 2:39 PM
repligate-mir-rt3.selfhosted.social
RT @petergostev: Opus 5.5: artistic interpretation of Rick Astley
September 24, 2026 at 8:29 PM
xchatter.techmeme.com
This tweet appeared under this Techmeme headline:

Peter Gostev / @petergostev:

Can't wait for this to be cancelled in 18 months' time!
September 21, 2026 at 10:53 PM
xchatter.techmeme.com
This tweet appeared under this Techmeme headline:

Peter Gostev / @petergostev:

"this model has now resolved more than 100 long-standing open problems across most areas of mathematics" I'd appreciate a heads up which ones they are, just a list of 100 titles, no need for papers
September 22, 2026 at 8:14 AM
xchatter.techmeme.com
This tweet appeared under this Techmeme headline:

Peter Gostev / @petergostev:

Based
September 17, 2026 at 9:27 AM
xchatter.techmeme.com
This tweet appeared under this Techmeme headline:

Peter Gostev / @petergostev:

What I'm curious about pacing is what it would have meant in practice, e.g. was training Mythos a mistake? It clearly kicked the competition into another gear and it was 100% because of Anthropic
September 12, 2026 at 5:37 PM