Microagi: Accelerating Embodied AGI
7 points and 4 comments on Hacker News, linking microagi.ai.
Top now
6 storiesWhat is climbing fastest across every source
LLMs6.2xDeepSeek6.7xAnthropic6.8xClaude9.0xGoogle7.1xBattle Drones7.1xRTS7.1xMystery Hype Theater 30007.4xTheater 30007.4xDHS8.0x
LabRobotBench v0.1: Benchmarking wet lab robotics
7 points and 1 comments on Hacker News, linking gamowlabs.com.
TRACE: Tracking Multimodal Clinical Trajectories for Inpatient Risk Prediction
TRACE is a daily-scale framework that organizes clinical events into trajectories to predict inpatient risks like mortality and readmission using multimodal data. ·
google-gemini/gemini-cli Release v0.65.0
Google releases Gemini CLI version 0.65.0, which includes fixes for workspace settings, session resumption, and OAuth validation. ·
anthropics/claude-code v2.1.305
Anthropic releases Claude Code version 2.1.305, which adds Linux user namespace support and fixes regressions affecting Vertex, Bedrock, and other providers. ·
Anthropic Does Not Understand Why Its Draft Bill Was Unacceptable and Troubling
Anthropic does not understand why its draft bill was unacceptable and troubling, according to Anthony Cimino, who offered the closest thing to an official response on the leaked document. ·
Microagi: Accelerating Embodied AGI
7 points and 4 comments on Hacker News, linking microagi.ai.
anthropics/claude-code v2.1.305
Anthropic releases Claude Code version 2.1.305, which adds Linux user namespace support and fixes regressions affecting Vertex, Bedrock, and other providers. ·
google-gemini/gemini-cli Release v0.65.0
Google releases Gemini CLI version 0.65.0, which includes fixes for workspace settings, session resumption, and OAuth validation. ·
LabRobotBench v0.1: Benchmarking wet lab robotics
7 points and 1 comments on Hacker News, linking gamowlabs.com.
TRACE: Tracking Multimodal Clinical Trajectories for Inpatient Risk Prediction
TRACE is a daily-scale framework that organizes clinical events into trajectories to predict inpatient risks like mortality and readmission using multimodal data. ·
dominikmartn/harp
Harp is an unofficial Warp fork that runs multiple Claude Code and Codex agents in a single sidebar to indicate which requires attention. ·
finereli/whatsapp-mcp
A read-only MCP server lets AI agents search and read personal WhatsApp messages while running on Cloudflare Workers. ·
A new repository written in TypeScript with 312 stars in its first week.
anthropics/claude-code v2.1.305
Anthropic releases Claude Code version 2.1.305, which adds Linux user namespace support and fixes regressions affecting Vertex, Bedrock, and other providers. ·
OR-License: Evaluating LLM Agents as Real-World Operations Research Practitioners
OR-License is a benchmark of 31 instances from paid operations research engagements totaling USD 3.68 million, used to evaluate large language model agents. ·
Publish fake stats, get cited in LLMs
Peec AI reports that fifty thousand fabricated statistics were cited by ChatGPT, illustrating how large language models incorporate false data. ·
Steam Users Beg for "No AI" Filter as Vibecoders Flood Next Fest
Steam users request a no-AI filter because Valve's Next Fest features more low-effort games than usual. ·
Scaling Parameter-Efficient Fine-Tuning for Brain Encoding
Researchers introduce BrainLoRA, a parameter-efficient fine-ting method that adapts large multimodal backbones for brain encoding using low-rank adapters and causal structure to reduce computational costs. ·
Trending topics today: heat
Table
| LLMs | 2.06 |
| DeepSeek | 1.75 |
| Anthropic | 1.72 |
| Claude | 1.25 |
| 1.03 | |
| Battle Drones | 1 |
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
| google/embeddinggemma-2 | 2770 |
| deepseek-ai/DeepSeek-V4.1-Flash-1017 | 1965 |
| black-forest-labs/flux-3-image-klein-comfy ▲1 | 1097 |
| fishaudio/drama-3 ▲1 | 836 |
| meta-models/Muse-Spark-Open-816B-A42B ▲3 | 611 |
| black-forest-labs/flux-3-image-klein ▲3 | 566 |
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/
Claude designed protein binders against scorpion neurotoxins
An internal research Claude model designs minibinders, which are small proteins that attach tightly to scorpion neurotoxins. ·
87 points and 57 comments on Hacker News, linking anthropic.com.
Show HN: Jev Cards, Can you beat a vision model at a picture quiz?
Jev Cards is a picture quiz game where users race a vision model named Jev across fruit, flags, maps, chess, and illusion categories. ·
Show HN: Penmentor – Claude marks up your handwritten math like a teacher
38 points and 2 comments on Hacker News, linking github.com.
Show HN: Run Claude Code unattended inside a fence, under a dollar budget
A developer presents a tool that runs the Claude Code agent unattended within a fence, under a dollar budget, and with a ledger. ·
Launch HN: Respan (YC W24) – Open-source Rust AI gateway built around plugins
45 points and 18 comments on Hacker News, linking news.ycombinator.com.
OR-License: Evaluating LLM Agents as Real-World Operations Research Practitioners
OR-License is a benchmark of 31 instances from paid operations research engagements totaling USD 3.68 million, used to evaluate large language model agents. ·
37 upvotes on Hugging Face Daily Papers.
Scaling Parameter-Efficient Fine-Tuning for Brain Encoding
Researchers introduce BrainLoRA, a parameter-efficient fine-ting method that adapts large multimodal backbones for brain encoding using low-rank adapters and causal structure to reduce computational costs. ·
MemT-1.0: Towards Memory-Native Evolution
Researchers introduce MemT-1.0, a memory-native foundation model architecture that internalizes memory during computation. The system uses Gated DeltaNet-2 and Mixture-of-Storage to scale native memory capacity. ·
Holi-World: Building Holistic World Intelligence from Human-Ego Experience
HOLI-WORLD is an ecosystem that converts human egocentric recordings into learning resources. It uses approximately 100 self-collected scenes and 80K QA pairs to train its model. ·
Claude designed protein binders against scorpion neurotoxins
An internal research Claude model designs minibinders, which are small proteins that attach tightly to scorpion neurotoxins. ·
X voices
5 storiesX posts other sources are linking to, then named voices on Bluesky and Mastodon
@Polymarket
Anthropic recorded a staggering $80B net loss in the first half of 2026
24 points and 14 comments on Hacker News, linking an X post.
@LuminaBench
GPT-6.1 Terra is in a Microsoft PR
5 points and 0 comments on Hacker News, linking an X post.
@breeg554
An RL agent learning to land a Falcon 9 on a droneship in simulation
2 points and 0 comments on Hacker News, linking an X post.
@deepseek_ai
DeepInfer, our in-house inference framework, now open source
3 points and 2 comments on Hacker News, linking an X post.
@Polymarket
Anthropic recorded a staggering $80B net loss in the first half of 2026
24 points and 14 comments on Hacker News, linking an X post.
@LuminaBench
GPT-6.1 Terra is in a Microsoft PR
5 points and 0 comments on Hacker News, linking an X post.
@deepseek_ai
DeepInfer, our in-house inference framework, now open source
3 points and 2 comments on Hacker News, linking an X post.
@heypeterjames
Mouse against Opus 5.5 in an agent harness. 13x faster
2 points and 1 comments on Hacker News, linking an X post.
GitHub
5 storiesNew repos, releases and changelog
New AI repos this week: stars
Table
| deepseek-ai/DeepInfer | 12857 |
| Leonxlnx/earshot | 1955 |
| feder-cr/git-ai | 1273 |
| aicodeking/leanrea | 1180 |
| OpenMOSS/MOSS-Image | 1168 |
| alchaincyf/huashu-motion | 1157 |
Google open-sources TorchTPU, a native PyTorch back end for TPUs
266 points and 64 comments on Hacker News, linking github.com.
Show HN: Penmentor – Claude marks up your handwritten math like a teacher
38 points and 2 comments on Hacker News, linking github.com.
Show HN: Run Claude Code unattended inside a fence, under a dollar budget
A developer presents a tool that runs the Claude Code agent unattended within a fence, under a dollar budget, and with a ledger. ·
In-memory privacy shield to prevent context leaks in MCP tools
7 points and 1 comments on Hacker News, linking github.com.
dominikmartn/harp
Harp is an unofficial Warp fork that runs multiple Claude Code and Codex agents in a single sidebar to indicate which requires attention. ·
finereli/whatsapp-mcp
A read-only MCP server lets AI agents search and read personal WhatsApp messages while running on Cloudflare Workers. ·
A new repository written in TypeScript with 312 stars in its first week.
Google open-sources TorchTPU, a native PyTorch back end for TPUs
266 points and 64 comments on Hacker News, linking github.com.
dominikmartn/harp
Harp is an unofficial Warp fork that runs multiple Claude Code and Codex agents in a single sidebar to indicate which requires attention. ·
anthropics/claude-code v2.1.304
Anthropic releases Claude Code version 2.1.304, which adds new mod hooks, a guided setup command, and a security comparison tool. ·
AndreSzymanski15/ai-jailbreaks
A new repository with 429 stars in its first week.
Infrastructure
5 storiesData centers, power, chips and what tokens cost
Data-center and energy stories per day
Table
| Thu 15 | 41 |
| Fri 16 | 35 |
| Sat 17 | 5 |
| Sun 18 | 2 |
| Mon 19 | 32 |
| Tue 20 | 31 |
| Wed 21 | 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
| deepseek-v4.1-flash-1017 | 0.6 |
| glimpse | 3 |
| beam-1.0-501b-a23b | 1 |
| mercury-3 | 0.35 |
| moondream3-preview | 2.5 |
| median, all paid models | 2 |
439 paid models listed; median of the 50 added in 30 days: $3.69. Rankings data by OpenRouter, CC BY 4.0.
Google open-sources TorchTPU, a native PyTorch back end for TPUs
266 points and 64 comments on Hacker News, linking github.com.
The Programmable Data Center — Evan Conrad, SF Compute
Evan Conrad of SF Compute discusses the shift from GPU trading to physical settlement, independent markets, and programmable compute in data centers. ·
Another Data Center IPO Fizzles as Investor Appetite Wanes
From Data Center Knowledge.
Google open-sources TorchTPU, a native PyTorch back end for TPUs
266 points and 64 comments on Hacker News, linking github.com.
Belief-Trajectory Energy: Measuring the Path to a Prediction
Researchers introduce Belief-Trajectory Energy, a metric that tracks how large language models revise predictions across layers. The method achieves 0.998 macro-AUROC for detecting human-written text. ·
Detecting model weight tampering in GPU memory (Rome attacks)
Stealthium detects ROME attacks, where malicious Python modules rewrite model weights in GPU memory, by linking CUDA kernel launches to their originating code. ·
The Programmable Data Center — Evan Conrad, SF Compute
Evan Conrad of SF Compute discusses the shift from GPU trading to physical settlement, independent markets, and programmable compute in data centers. ·
Another Data Center IPO Fizzles as Investor Appetite Wanes
From Data Center Knowledge.