Intel stock collapse worsens as $3.5T chip rout spreads
Intel stock declines further as a $3.5 trillion chip sector sell-off spreads to memory and storage companies on Monday. ·
68 points and 36 comments on Hacker News, linking finance.yahoo.com.
Top now
6 storiesWhat is climbing fastest across every source
Anthropic6.8xOpenAI9.9xS-18.1xAstra7.5xIPO6.7xRecursive Self Improvement7.8xCensus Bureau7.5xBalanov8.6xChina7.6xAgents6.2x
The SEC Should Block Anthropic's IPO to Protect the Public
Matt Scherer writes that the SEC should block Anthropic's IPO to protect workers, retail investors, and retirees from future bills. ·
llm 0.37
llm 0.37 updates the LLM CLI tool with new models, an OpenAI endpoint command, and embedding support for Gemini Embedding 2. ·
Building AI for Lawyers - Jacob Lauritzen, CTO of Legora
From Latent Space.
Building Legal Agents When You Can't See the Data - Jacob Lauritzen, CTO of Legora
Jacob Lauritzen, CTO of Legora, explains how the company builds and tests legal agents without seeing client data traces. ·
sgl-project/sglang v0.5.23
SGLang version 0.5.23 includes 1007 pull requests from 257 contributors and adds support for models like Muse Spark Open and Beam-1.0. ·
Intel stock collapse worsens as $3.5T chip rout spreads
Intel stock declines further as a $3.5 trillion chip sector sell-off spreads to memory and storage companies on Monday. ·
68 points and 36 comments on Hacker News, linking finance.yahoo.com.
Anthropic's growth just flattened
162 points and 190 comments on Hacker News, linking regenerator1.com.
Balanov/shim
Balanov/shim is a tiny local-first AI harness supporting any model, MCP, and Agent Skills via a natural-language CLI with zero dependencies under the MIT license. ·
langchain==1.4.8
LangChain version 1.4.8 updates dependencies and warns that Python 3.10 support ends in version 1.5.0. ·
vllm-project/vllm v0.33.0
vLLM version 0.33.0 adds support for 122 new models and upgrades to PyTorch 2.14.0 while dropping CUDA 12.9. ·
FusionCow/Sulphur-3
FusionCow releases the Sulphur-3 model, a minimax-h3 machine learning model available for download and exploration on Hugging Face. ·
burner22/numbat
The burner22/numbat model is a machine learning model with 6.8 trillion parameters available on Hugging Face. ·
A model model from burner22 with 298 likes and 7,108 downloads on Hugging Face.
Astra can read and write weights to improve Qwen
Astra reads and writes weights to improve Qwen. Frontier LLMs natively read and write neural network weights better than expected. ·
Anthropic's growth just flattened
162 points and 190 comments on Hacker News, linking regenerator1.com.
vllm-project/vllm v0.33.0
vLLM version 0.33.0 adds support for 122 new models and upgrades to PyTorch 2.14.0 while dropping CUDA 12.9. ·
No Channels Left Behind: Understanding Diffusion Acceleration through Residual Gates
Researchers propose that diffusion transformer acceleration stems from reshaping AdaLN-Zero residual gates, introducing a teacher-free radial objective to improve generation quality. ·
FrameForge: Harnessing Priors of Video Foundation Models for Unified Image Generation
FrameForge uses video foundation models to unify conditional image generation through mixture of temporal experts and chain of temporal reasoning. ·
Trending topics today: heat
Table
| Anthropic | 4.73 |
| OpenAI | 2.1 |
| S-1 | 1.86 |
| Astra | 1.58 |
| IPO | 1.22 |
| Recursive Self Improvement | 1.09 |
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
| deepseek-ai/DeepSeek-V4.1-Flash-1017 ▼1 | 4704 |
| mistralai/Mistral-Large-4-1T-A52B | 1501 |
| LiquidAI/LFM3-1.7B ▲1 | 1069 |
| nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B | 684 |
| baidu/Unlimited-OCR-2 ▲1 | 563 |
| speridlabs/idol ▲4 | 419 |
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/
Astra can read and write weights to improve Qwen
Astra reads and writes weights to improve Qwen. Frontier LLMs natively read and write neural network weights better than expected. ·
1,945 points and 619 comments on Hacker News, linking rikiyat.github.io.
Launch HN: Momentic (YC W24) – Agents that catch bugs by using your app
Momentic is an AI QA engineer that catches bugs by using your app. It has caught 117,010 bugs with repro steps and evidence. ·
Claude Opus 5.5 Can Draw Permit-Level Architectural Drawings
Claude Opus 5.5 generates permit-level architectural drawings, with testing showing its drafting capabilities approach professional standards. ·
FusionCow/Sulphur-3
FusionCow releases the Sulphur-3 model, a minimax-h3 machine learning model available for download and exploration on Hugging Face. ·
FrameForge: Harnessing Priors of Video Foundation Models for Unified Image Generation
FrameForge uses video foundation models to unify conditional image generation through mixture of temporal experts and chain of temporal reasoning. ·
burner22/numbat
The burner22/numbat model is a machine learning model with 6.8 trillion parameters available on Hugging Face. ·
A model model from burner22 with 298 likes and 7,108 downloads on Hugging Face.
Astra can read and write weights to improve Qwen
Astra reads and writes weights to improve Qwen. Frontier LLMs natively read and write neural network weights better than expected. ·
No Channels Left Behind: Understanding Diffusion Acceleration through Residual Gates
Researchers propose that diffusion transformer acceleration stems from reshaping AdaLN-Zero residual gates, introducing a teacher-free radial objective to improve generation quality. ·
FrameForge: Harnessing Priors of Video Foundation Models for Unified Image Generation
FrameForge uses video foundation models to unify conditional image generation through mixture of temporal experts and chain of temporal reasoning. ·
Launch HN: Momentic (YC W24) – Agents that catch bugs by using your app
Momentic is an AI QA engineer that catches bugs by using your app. It has caught 117,010 bugs with repro steps and evidence. ·
X voices
5 storiesX posts other sources are linking to, then named voices on Bluesky and Mastodon
@OpenAIDevs
Pando - OpenAI's alternative to Git for source control
1 points and 0 comments on Hacker News, linking an X post.
@Prosaic_int
AI models are enabling suicide
4 points and 3 comments on Hacker News, linking an X post.
@TheVixhal
QwenGPU: A functioning, deterministic SIMT GPU inside Qwen3.5
4 points and 0 comments on Hacker News, linking an X post.
@huskirl
Asking my AI teacher how to spell seventeen
12 points and 4 comments on Hacker News, linking an X post.
@wallstengine
Amazon plans to start selling Trainium AI chips outside AWS
3 points and 0 comments on Hacker News, linking an X post.
@jiahaoshao1
Anthropic blocks users for anything touching training larger models
488 points and 266 comments on Hacker News, linking an X post.
@huskirl
Asking my AI teacher how to spell seventeen
12 points and 4 comments on Hacker News, linking an X post.
@Prosaic_int
AI models are enabling suicide
4 points and 3 comments on Hacker News, linking an X post.
@TheVixhal
QwenGPU: A functioning, deterministic SIMT GPU inside Qwen3.5
4 points and 0 comments on Hacker News, linking an X post.
@motatoeshq
Background agents are dead, long live interactive agents
2 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
| deepseek-ai/dsh-superagent-prototype | 13235 |
| heydamir/unslop | 6286 |
| wangzan101/Mika | 5317 |
| paipiange/nvzhuang | 2043 |
| caru-ini/stack-diagram | 1203 |
| jacke-pp/open-yoki | 841 |
Balanov/shim
Balanov/shim is a tiny local-first AI harness supporting any model, MCP, and Agent Skills via a natural-language CLI with zero dependencies under the MIT license. ·
A new repository written in Go with 170 stars in its first week.
langchain==1.4.8
LangChain version 1.4.8 updates dependencies and warns that Python 3.10 support ends in version 1.5.0. ·
vllm-project/vllm v0.33.0
vLLM version 0.33.0 adds support for 122 new models and upgrades to PyTorch 2.14.0 while dropping CUDA 12.9. ·
anthropics/anthropic-sdk-python v2.4.0
Anthropic releases version 2.4.0 of its Python SDK, adding GA methods for analytics, spend limits, and RBAC. ·
Balanov/uncut
A new repository written in Go with 572 stars in its first week.
vllm-project/vllm v0.33.0
vLLM version 0.33.0 adds support for 122 new models and upgrades to PyTorch 2.14.0 while dropping CUDA 12.9. ·
New release v0.33.0rc3 of vllm-project/vllm on GitHub.
Show HN: A working Nintendo in ~1GB of Qwen weights
294 points and 152 comments on Hacker News, linking github.com.
Balanov/uncut
A new repository written in Go with 572 stars in its first week.
davidlopez-x05x8/minecraft-aim-assist
Davidlopez-x05x8 releases Minecraft Aim Assist 2026, a Fabric mod for version 1.21.4 that provides smooth aiming assistance with adaptive tracking and customizable FOV. ·
caru-ini/stack-diagram
The caru-ini/stack-diagram agent skill generates an architecture diagram with official logos from a single sentence describing a technology stack. ·
Infrastructure
5 storiesData centers, power, chips and what tokens cost
Data-center and energy stories per day
Table
| Wed 28 | 32 |
| Thu 29 | 42 |
| Fri 30 | 29 |
| Sat 31 | 2 |
| Sun 01 | 5 |
| Mon 02 | 34 |
| Tue 03 | 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
| glm-5.3-flashx-legal | 1.25 |
| particle-1 | 0.5 |
| gpt-6.1-terra-pro | 5 |
| gpt-6.1-terra-pro:batch | 2.5 |
| gpt-6.1-terra | 5 |
| median, all paid models | 2 |
425 paid models listed; median of the 29 added in 30 days: $2.50. Rankings data by OpenRouter, CC BY 4.0.
SpaceXAI launching two additional data center campuses in Mississippi to further expand Colossus 2
SpaceXAI launches two additional data center campuses in Mississippi to expand Colossus 2 Sites, dubbed Gun Range and Spam, which will be operational by the end of the year. ·
Liquid Cooling Retrofits: Hard, but Sometimes Worth It
Liquid cooling retrofits increase density and efficiency, but cost, downtime, and infrastructure limits restrict their use to select cases. ·
America on Data Centers: listen to 10,000+ city council comments
The site displays over ten thousand public comments on data centers from local government meetings across forty-eight states. ·
14 points and 22 comments on Hacker News, linking americansentiment.com.
Sam Altman says DRAM capacity should go to data centers, not personal computers
OpenAI CEO Sam Altman says DRAM capacity should go to data centers, not personal computers. ·
Hermantown approves controversial Google Data Center during late night meeting
The Hermantown City Council unanimously approves six measures allowing a Google data center to proceed near Duluth, and the proposal still faces three lawsuits. ·