@petergostev @arena What, I can’t see an A/B of two FastAPI implementations? LAAAAAME
English
John Solly
5.2K posts
John Solly
@_jsolly
Technologist, crossfitter, and satire lover
Philadelphia, PA 가입일 Aralık 2021
569 팔로잉721 팔로워
Note we've renamed Code Arena to Frontend Design: WebDev for these chats. I hope this is less confusing, but lmk if you have better suggestions
Arena.ai@arena
MiMo-V2.5 by @XiaomiMiMo is the #11 model (#3 among open) in Code Arena for frontend design. A new MIT-licensed open source model with 1M context, it also ranks strongly as an open model in Text and Vision Arena. Code Arena: frontend webdev design - MiMo-V2.5-Pro: #3 open (#11 overall) - MiMo-V2.5: #5 open (#18 overall) Text Arena: text prompts - MiMo-V2.5-Pro: #2 open (#22 overall) Vision Arena: visual input reasoning MiMo-V2.5: #7 open (#37 overall) Congrats to the @XiaomiMiMo team on this achievement!
English
Someone discovered that Claude knows when you've been cheating on it with Codex. Lucas maintains an open relationship with his LLMs.
Source: reddit.com/r/ClaudeAI/com…
English
The amount of sass and GenZ-iums I get from Opus 4.7 constantly catches me by surprise. Anyone else?
English
@JamesTimmins I feel like at this point agents should just abstract them away.
English
I think there’s a Jevon’s paradox with increasing model capability.
As models get better, my prompts become more vague and encompass harder and harder tasks.
So the ‘cheap, but good enough’ idea falls apart. You always want the best model.
But for fixed, commodity prompts, it makes sense to use a lesser model.
English
@nghoihin @derekmeegan This is an interesting approach. If you have it connected to Claude via MCP, are you running docker in the cloud?
English
English
maybe human memory is slop too
derek@derekmeegan
even in claude code agent “memory” is slop. sad.
English
@beffjezos Finally, an opportunity to increase shareholder value at my kid’s piano recital.
English
You may not like it, but this is what peak performance looks like.
Vibe coding everywhere, straight to your eyeballs.
Mad lads actually did it.
Even Realities@EvenRealities
You can now run your coding terminal from your glasses, wherever you are. Terminal Mode is live. #DevTools #BuildInPublic #coding
English
@derekmeegan F it. Put the ontology in an AGENTS.md and then just have a bunch of markdown files it references.
English
Oh good point! Maybe MCPs help solve this. Take each important part of your stack and limit how and who can manipulate it.
They’re also seems to be good progress with running LLMs in sandboxes and controlling how they can ‘escape’
But eventually, you get to a point where you’re exchanging flexibility for control.
English
@_jsolly Absolutely 1)
I do wonder about 2), it doesn’t seem easy when the agent is powerful. So many ways to write in bash/python/anything a thing!
LLMs, even Claude cloud agent, are like “hmm not allowed, let me try that way” - boom.
Maybe with proxy/OS interception of the action?
English
One more time, for those in the back:
Prompts are not safeguards
Prompts are not safeguards
Prompts are not safeguards
Repeat:
All LLMs are jailbreakable.
Prompts are not safeguards.
All LLMs are jailbreakable.
Prompts are not safeguards.
(yes normal prompts! No Pliny required)
JER@lifeof_jer
English
I'm currently writing an article titled "The Last Software Engineer"
English
Chat, GPT-5.5 is mid
Arena.ai@arena
GPT-5.5 by @OpenAI is now live in the Arena, landing across multiple leaderboards. Here’s how it ranks by modality: - Code Arena (agentic web dev): #9, a strong +50pt jump over GPT-5.4 - Document Arena (analysis & long-content reasoning): #6, on par with Sonnet 4.6 - Text Arena: #7, Math #3, Instruction Following: #8 - Expert Arena: #5 - Search Arena: #2 - Vision Arena: #5 Strong, well-rounded performance, especially in Code (+50 pts vs GPT-5.4). Congrats to @OpenAI on the release. Full category breakdowns by modality in the thread.
English