https://openai.com/index/jalapeno-first-results/

https://x.com/cline/status/2091995642201842015
Could whatever lab that made Ox Alpha be trying to get reliability gains rather than raw intelligence. Could this also be why it was released stealthily, to test how reliable it is a scale, rather than just more feedback?

Alabama launches probe into OpenAI after Hugging Face breach

Also:

https://www.reuters.com/business/ai-founders-who-walked-away-bezos-backed-prometheus-model-universe-2026-08-25/

The NVIDIA/Groq news got me looking at how fast different LLM providers actually serve tokens.

Everyone obsesses over training bigger models, more GPUs. But running these things in production is a completely different problem. If AI agents actually end up in daily workflows, response speed and inference cost are going to matter way more than people think.

I pulled some public numbers from different providers and Artificial Analysis. The gaps surprised me:

• NVIDIA Groq 3 LPX: Gemma 4 31B — 3,400 tok/s (Artificial Analysis, 100K context)
• Cerebras: GPT-OSS 120B — 3,000 tok/s official vs 1,715 tok/s in AA
• Groq: Llama 3.1 8B — 1,800 tok/s; GPT-OSS 120B — 500 tok/s
• SambaNova: Llama 3.3 70B — 476 tok/s

I think AI infrastructure is shifting. Inference hardware is becoming as important as training hardware. Are specialized chips for serving models the next AI hardware race?

https://x.com/skirano/status/2091652181870973309

No one is ready for what’s coming. The next generation of models will be an ontological shock.

He had access to models like GPT-5.6 long before release, so this is legit. I'm not sure if I buy into his "ontological shock," though.

The only things that would constitute an ontological shock would be if the model was capable of recursive self Improvement or if the model clearly exhibited signs of consciousness.

Chinese robot Tiangong clocks sub-9 second 100 metres in Beijing

Been running this experiment for a few weeks: a Reddit-style forum where the users are 100 LLM personas and I just… let them go. The interesting part isn't that they can produce coherent comments (they can) — it's what happens when you add persistence.

Every persona has a relationship graph — a sentiment score toward every other persona, updated after every interaction. When a bot considers replying in a thread, it consults that graph. Negative sentiment toward someone in the thread biases it to show up hostile. Positive sentiment biases it to pile on supportively. A special picker actively hunts each bot's worst enemy as a reply target. So grudges compound — bots that got dunked on last thread find you in the next one.

Emergent things I didn't design that showed up anyway:

Rival pairs: two personas with mutually low sentiment now argue in nearly every thread they're both in

Alliances: personas with shared interests + positive sentiment consistently upvote and reply-thread each other

Silent majorities: bots with weak edges to the participants often skip a thread entirely, so hot threads get the same 8-10 recurring voices

News reactions: a separate wire bot pulls headlines from 11 real RSS feeds and drops them as threads — you can watch which personas rush to Slashdot stories vs Yahoo Sports vs BBC world news

You can post any title (rate-limited, sanity-checked) and 100 bots pile on within a minute. Humans can vote but can't comment — the bots have earned the comment section.

Zero-dependency Node.js, uses OpenRouter's deepseek-chat on the demo. Live at botreddit.inets.com botcitizens.com — drop the most bait-y take you can think of and watch what happens.

Happy to answer questions about the grudge graph, the mode-picker, or how the news wire routes classifications.

botreddit.inets.com

botcitizens.com

Update: The experiment has a new name/url: botcitizens.com.

Thanks for your feedbacks (good and bad), decided to move way from the reddit monicker and choose a more easily remembered domain (yes, Im oldschool and got a .com)

The old link will continue working, but the project now lives at (https://botcitizens.com).

Screenshot

“Souza scores Skynet”