Quoting Ben Affleck
Ben Affleck explains how convolutional neural networks process pixel data to identify visual patterns for green screen replacement in film production. ·
From Simon Willison.
Top now
6 storiesWhat is climbing fastest across every source
Claude Haiku 5.57.0xClaude6.8xOpenAI5.7xOct7.0xGPT-67.0xLLMs6.1xMicrosoft7.0xChatGPT4.7xMeta5.9xMistral Large 44.5x
Claude Haiku 5.5
627 points and 316 comments on Hacker News, linking anthropic.com.
Meta and Microsoft take steps to reduce employee usage of Claude AI
Meta and Microsoft implement measures to limit employee use of Claude AI. ·
anthropics/claude-code v2.1.293
Anthropic releases Claude Haiku 5.5 as the default Haiku model on its API, featuring a 1M token context window and pricing of $0.10 per million input tokens. ·
langchain-fireworks==1.7.1
LangChain releases version 1.7.1 of its Fireworks integration, which preserves reasoning during streaming and tool loops while updating the LangGraph SDK dependency. ·
Does better work always mean better workers?
From Google Research.
Claude Haiku 5.5
627 points and 316 comments on Hacker News, linking anthropic.com.
Meta and Microsoft take steps to reduce employee usage of Claude AI
Meta and Microsoft implement measures to limit employee use of Claude AI. ·
anthropics/claude-code v2.1.293
Anthropic releases Claude Haiku 5.5 as the default Haiku model on its API, featuring a 1M token context window and pricing of $0.10 per million input tokens. ·
unslothai/unsloth Train your own Decision model
Unsloth releases a tool that converts text or vision LLMs into decision models, raising accuracy from 30% to 80%. ·
canberkkkkkk/ema-lightning
A text to speech model from canberkkkkkk with 263 likes and 2,724 downloads on Hugging Face.
unsloth/embeddinggemma-2-GGUF
A feature extraction model from unsloth with 161 likes and 11.5k downloads on Hugging Face.
Claude Haiku 5.5
627 points and 316 comments on Hacker News, linking anthropic.com.
Mistral Large 4: "Le Chonk"
Mistral Large 4, nicknamed Le Chonk, is an enterprise AI platform that allows users to customize, fine-tune, and deploy assistants, agents, and multimodal models. ·
anthropics/claude-code v2.1.293
Anthropic releases Claude Haiku 5.5 as the default Haiku model on its API, featuring a 1M token context window and pricing of $0.10 per million input tokens. ·
unslothai/unsloth Train your own Decision model
Unsloth releases a tool that converts text or vision LLMs into decision models, raising accuracy from 30% to 80%. ·
canberkkkkkk/ema-lightning
A text to speech model from canberkkkkkk with 263 likes and 2,724 downloads on Hugging Face.
Rethinking Cross-Tokenizer On-Policy Distillation: From Alignment Coverage to Supervision Reliability
Researchers find that restricting reverse KL to a student-selected top-16 vocabulary subset achieves accuracy comparable to full shared-vocabulary on-policy distillation, outperforming evaluated cross-tokenizer baselines. ·
Trending topics today: heat
Table
| Claude Haiku 5.5 | 2.57 |
| Claude | 2.15 |
| OpenAI | 2.09 |
| Oct | 1.22 |
| GPT-6 | 1.15 |
| LLMs | 1.03 |
Heat = summed hotness of today's stories naming the topic.
Models
5 storiesReleases, papers, trending weights and new listings
Top 6 trending on Hugging Face: likes
Table
| abenzerps/Qwen-Image-2.1-Uncensored-GGUF ▼3 | 3558 |
| autotrust/JEV-27B-VL | 1916 |
| Cloudflare/clef | 1813 |
| autotrust/GEV-26B-Decide ▲4 | 1273 |
| google/embeddinggemma-2 ▲6 | 941 |
| Aleph-Alpha/Kolibri-1 ▼1 | 774 |
PirateBench: score out of 100
Table
| gpt-6.1-sol | 98.9 |
| opus-5-5 | 96.46 |
| gpt-6-sol | 91.6 |
| gpt-6-luna | 65.25 |
| haiku-4-5 | 4.28 |
Agents build our own projects from a brief; every run, failures included: https://cowerx.dev/daily/piratebench/
unslothai/unsloth Train your own Decision model
Unsloth releases a tool that converts text or vision LLMs into decision models, raising accuracy from 30% to 80%. ·
New release v0.1.904-beta of unslothai/unsloth on GitHub.
AI Model Groupthink
15 points and 3 comments on Hacker News, linking magicnumbers.io.
Tell HN: Apple not letting removal of AI models on macOS 27 is outrageous
8 points and 0 comments on Hacker News, linking news.ycombinator.com.
canberkkkkkk/ema-lightning
A text to speech model from canberkkkkkk with 263 likes and 2,724 downloads on Hugging Face.
Mistral Large 4: "Le Chonk"
Mistral Large 4, nicknamed Le Chonk, is an enterprise AI platform that allows users to customize, fine-tune, and deploy assistants, agents, and multimodal models. ·
Mistral Large 4: "Le Chonk"
Mistral Large 4, nicknamed Le Chonk, is an enterprise AI platform that allows users to customize, fine-tune, and deploy assistants, agents, and multimodal models. ·
521 points and 5 comments on Hacker News, linking mistral.ai.
EmbeddingGemma 2: An open, lightweight multimodal embedding model
413 points and 46 comments on Hacker News, linking blog.google.
unslothai/unsloth Train your own Decision model
Unsloth releases a tool that converts text or vision LLMs into decision models, raising accuracy from 30% to 80%. ·
canberkkkkkk/ema-lightning
A text to speech model from canberkkkkkk with 263 likes and 2,724 downloads on Hugging Face.
Rethinking Cross-Tokenizer On-Policy Distillation: From Alignment Coverage to Supervision Reliability
Researchers find that restricting reverse KL to a student-selected top-16 vocabulary subset achieves accuracy comparable to full shared-vocabulary on-policy distillation, outperforming evaluated cross-tokenizer baselines. ·
X voices
5 storiesX posts other sources are linking to, then named voices on Bluesky and Mastodon
@davepl1968
Dave W Plummer: "AI Cloning should be illegal"
4 points and 6 comments on Hacker News, linking an X post.
@p_rabtsevich
I found an exoplanet using Claude, and now you can fly to it yourself
2 points and 0 comments on Hacker News, linking an X post.
@elonmusk
Grok Bot will use Opus 5.5, Midjourney
3 points and 0 comments on Hacker News, linking an X post.
@github
Building Git infrastructure for agent-scale development
3 points and 0 comments on Hacker News, linking an X post.
@ShaneMac
AI agent posted personal bank balances into company Slack
2 points and 1 comments on Hacker News, linking an X post.
@davepl1968
Dave W Plummer: "AI Cloning should be illegal"
4 points and 6 comments on Hacker News, linking an X post.
@p_rabtsevich
I found an exoplanet using Claude, and now you can fly to it yourself
2 points and 0 comments on Hacker News, linking an X post.
@elonmusk
Grok Bot will use Opus 5.5, Midjourney
3 points and 0 comments on Hacker News, linking an X post.
@github
Building Git infrastructure for agent-scale development
3 points and 0 comments on Hacker News, linking an X post.
GitHub
5 storiesNew repos, releases and changelog
New AI repos this week: stars
Table
| QingYunA/answer-me-with-html | 2036 |
| Jakeschincariol/replica-skill | 851 |
| StayLameBro/backburner | 844 |
| elstongun/leviathan | 665 |
| AgentMemoryRepo/agentmemoryrepo | 613 |
| DozenTwelve/Papermorph | 529 |
Docker Agent
163 points and 75 comments on Hacker News, linking github.com.
unslothai/unsloth Train your own Decision model
Unsloth releases a tool that converts text or vision LLMs into decision models, raising accuracy from 30% to 80%. ·
openai/codex 0.161.0
OpenAI releases Codex 0.161.0, which makes GPT-6.1 Sol the default model in bundled and Amazon Bedrock catalogs. ·
AI-assisted proof of optimal packing for 11 squares
Researchers use AI assistance to formally verify the optimal packing arrangement for eleven squares within a larger square. ·
Show HN: Pinrail – A desktop inbox where coding agents wait for your review
3 points and 0 comments on Hacker News, linking github.com.
unslothai/unsloth Train your own Decision model
Unsloth releases a tool that converts text or vision LLMs into decision models, raising accuracy from 30% to 80%. ·
New release v0.1.904-beta of unslothai/unsloth on GitHub.
Docker Agent
163 points and 75 comments on Hacker News, linking github.com.
AI-assisted proof of optimal packing for 11 squares
Researchers use AI assistance to formally verify the optimal packing arrangement for eleven squares within a larger square. ·
openai/codex 0.161.0
OpenAI releases Codex 0.161.0, which makes GPT-6.1 Sol the default model in bundled and Amazon Bedrock catalogs. ·
LosaLosSantos/aurelio-finance
LosaLosSantos releases aurelio-finance, a self-hosted tool for tracking personal wealth and portfolios using AI insights, built with FastAPI and React. ·
Infrastructure
5 storiesData centers, power, chips and what tokens cost
Data-center and energy stories per day
Table
| Fri 02 | 32 |
| Sat 03 | 0 |
| Sun 04 | 3 |
| Mon 05 | 31 |
| Tue 06 | 27 |
| Wed 07 | 28 |
| Thu 08 | 0 |
Stories from DCD, Data Center Knowledge, EIA and Hacker News, by publication day.
Median output price of models added in the last 30 days, $ per million tokens
One point per OpenRouter reading, last 8 days. Data by OpenRouter, CC BY 4.0.
Output price, $ per million tokens (newest models)
Table
| claude-haiku-5.5 | 0.5 |
| claude-haiku-5.5:batch | 0.25 |
| gemini-nano-banana-2.1 | 7.5 |
| mistral-large-4-0 | 2.09 |
| pareto-26.10-preview | 3.2 |
| median, all paid models | 2 |
440 paid models listed; median of the 51 added in 30 days: $2.09. Rankings data by OpenRouter, CC BY 4.0.
How many GPUs is 1M/B/T tokens?
A calculator converts daily, monthly, or yearly token volumes into estimated counts of A100, H100, or B200 GPUs, including a sourced uncertainty range. ·
14 points and 4 comments on Hacker News, linking cedana.com.
Waterless fire suppression is becoming key for data center community relations
Data centers increasingly use waterless fire suppression systems to manage community relations, as the risk of fire rises with the growth in facility quantity and size. ·
OpenTPU – An open-source AI accelerator, developed by AI
334 points and 391 comments on Hacker News, linking github.com.
OpenTPU – An open-source AI accelerator, developed by AI
334 points and 391 comments on Hacker News, linking github.com.
How many GPUs is 1M/B/T tokens?
A calculator converts daily, monthly, or yearly token volumes into estimated counts of A100, H100, or B200 GPUs, including a sourced uncertainty range. ·
Improper redaction reveals Google Data Center water and electricity usage
Nebraska data centers must file annual reports with the state water and energy department, but statutes prevent public access to their electricity and water usage figures. ·
Waterless fire suppression is becoming key for data center community relations
Data centers increasingly use waterless fire suppression systems to manage community relations, as the risk of fire rises with the growth in facility quantity and size. ·