A year after the 100% AI mandate nobody understands our code
36 points and 31 comments on Hacker News, linking reddit.com.
Top now
6 storiesWhat is climbing fastest across every source
Fable 5.27.5xClaude Fable 5.28.6xNVIDIA9.3xMuse Ember8.0xOpenRouter7.2xControl Room8.5xAnthropic7.3xOpenAI8.0xTPUs8.4xChinese6.9x
TPUs
127 points and 17 comments on Hacker News, linking blog.google.
New on OpenRouter: Thinking Machines: Inkling 2 Small (free)
Now on OpenRouter at $0 in and $0 out per million tokens.
anthropics/claude-code v2.1.321
Anthropic releases Claude Code v2.1.321, which adds Claude Fable 5.2 as the default model and updates remote-control settings. ·
langchain-fireworks==1.7.3
LangChain Fireworks version 1.7.3 updates default max tokens from model profiles and refreshes model profile data. ·
Exploring the world's languages with Language Explorer
From Google Research.
A year after the 100% AI mandate nobody understands our code
36 points and 31 comments on Hacker News, linking reddit.com.
TPUs
127 points and 17 comments on Hacker News, linking blog.google.
anthropics/claude-code v2.1.321
Anthropic releases Claude Code v2.1.321, which adds Claude Fable 5.2 as the default model and updates remote-control settings. ·
tencent/Qwen3.8-27B-DSpine-b16
Tencent releases Qwen3.8-27B-DSpine-b16, a text generation model with 2 billion parameters available on Hugging Face. ·
VideoRebirthAI/hyperflow-h3-flash
Hyperflow H3 Flash is a 14B parameter image-text-to-video model by Minimax available on Hugging Face. ·
IWIBench: Evaluating Gradient-Free Model Driven Intra-weight Injection
IWIBench evaluates gradient-free intra-weight injection by measuring how well editor models modify target model weights to achieve text-editing objectives. ·
Claude Fable 5.2 and Mythos 5.2
850 points and 514 comments on Hacker News, linking anthropic.com.
anthropics/claude-code v2.1.321
Anthropic releases Claude Code v2.1.321, which adds Claude Fable 5.2 as the default model and updates remote-control settings. ·
Claude Fable 5.2 Scores 65 on Artificial Analysis, Highest So Far
Anthropic's Claude Fable 5.2 achieves a score of 65 on the Artificial Analysis benchmark, marking its highest result to date. ·
SplatAnything: Gaussian Spacing Normalization for Feed-forward 3DGS across Arbitrary Scenes
60 upvotes on Hugging Face Daily Papers.
PCBRouteAgent: How Close Are LLM Agents to Becoming PCB Routing Engineers?
PCBRouteAgent automates printed circuit board routing using a multimodal agent that coordinates global planning with local execution via KiCad. ·
Scantist/GLM-5.3-SLIM-GGUF
Scantist releases GLM-5.3-SLIM-GGUF, a text generation model with 563 billion parameters available on Hugging Face. ·
Trending topics today: heat
Table
| Fable 5.2 | 2.82 |
| Claude Fable 5.2 | 2.59 |
| NVIDIA | 2.01 |
| Muse Ember | 1.79 |
| OpenRouter | 1.39 |
| Control Room | 1.34 |
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
| Lightricks/LTX-2.5 ▲1 | 11261 |
| moonshotai/Kimi-K3.1-Pro | 1796 |
| speridlabs/idol | 1625 |
| xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic | 1037 |
| ccore/Qwen3.5-4B-3D-Coder ▼2 | 949 |
| ricursiveAI/Ripple-65B-Preview ▲1 | 453 |
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 Fable 5.2 Scores 65 on Artificial Analysis, Highest So Far
Anthropic's Claude Fable 5.2 achieves a score of 65 on the Artificial Analysis benchmark, marking its highest result to date. ·
253 points and 167 comments on Hacker News, linking artificialanalysis.ai.
Claude Fable 5.2 and Mythos 5.2
850 points and 514 comments on Hacker News, linking anthropic.com.
Google Gemini routes personal Drive files and emails to human annotators
307 points and 62 comments on Hacker News, linking github.com.
Our Smallest, Most Efficient Model: Muse Ember
225 points and 93 comments on Hacker News, linking research.meta.ai.
Claude Fable 5.2 launch offers
8 points and 1 comments on Hacker News, linking support.claude.com.
Claude Fable 5.2 and Mythos 5.2
850 points and 514 comments on Hacker News, linking anthropic.com.
Claude Fable 5.2 Scores 65 on Artificial Analysis, Highest So Far
Anthropic's Claude Fable 5.2 achieves a score of 65 on the Artificial Analysis benchmark, marking its highest result to date. ·
Google Gemini routes personal Drive files and emails to human annotators
307 points and 62 comments on Hacker News, linking github.com.
Our Smallest, Most Efficient Model: Muse Ember
225 points and 93 comments on Hacker News, linking research.meta.ai.
SplatAnything: Gaussian Spacing Normalization for Feed-forward 3DGS across Arbitrary Scenes
60 upvotes on Hugging Face Daily Papers.
X voices
5 storiesX posts other sources are linking to, then named voices on Bluesky and Mastodon
@jalcowastaken
Show HN: Drop, self hosted, OSS file transfer for agents
1 points and 0 comments on Hacker News, linking an X post.
@ccheever
Exact: A new agent-native app framework written in Rust, from Expo team
3 points and 0 comments on Hacker News, linking an X post.
@CulperResearch
We are short Nvidia $NVDA. We believe Nvidia has a significant China problem
5 points and 0 comments on Hacker News, linking an X post.
@AgentMemoryL
AML Cycle 2 results are approaching: What should agent memory benchmarks reveal?
4 points and 0 comments on Hacker News, linking an X post.
@CulperResearch
We are short Nvidia $NVDA. We believe Nvidia has a significant China problem
5 points and 0 comments on Hacker News, linking an X post.
@ccheever
Exact: A new agent-native app framework written in Rust, from Expo team
3 points and 0 comments on Hacker News, linking an X post.
@0xAlbus
Grok bot helped a guy disprove a stalking charge in divorce court
3 points and 0 comments on Hacker News, linking an X post.
@AgentMemoryL
AML Cycle 2 results are approaching: What should agent memory benchmarks reveal?
4 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
| KKKKhazix/agent-autoloop | 4270 |
| janestreet/agent_skills | 2001 |
| sarpstar/cinecode | 1910 |
| Sandflow-ai/sand3d | 1613 |
| shanselman/gputaskman | 1475 |
| DIYgod/pinky | 1419 |
Google Gemini routes personal Drive files and emails to human annotators
307 points and 62 comments on Hacker News, linking github.com.
Show HN: STE80 Humanizer – de-AI text by writing 80% of the way to ASD-STE100
39 points and 11 comments on Hacker News, linking github.com.
The same H100 costs $2.69 or $12.29 an hour depending where you rent it
164 points and 84 comments on Hacker News, linking github.com.
6 Hours of YouTube and 85 handtyped pages of notes on LLM Inference – Start here
Samir Sarwar releases samir-vllm, a small LLM inference engine built from scratch with six hours of YouTube tutorials and eighty-five pages of handwritten notes. ·
Show HN: Flide: Make Jev-like decision models up to 9.5x faster
99 points and 24 comments on Hacker News, linking github.com.
Google Gemini routes personal Drive files and emails to human annotators
307 points and 62 comments on Hacker News, linking github.com.
The same H100 costs $2.69 or $12.29 an hour depending where you rent it
164 points and 84 comments on Hacker News, linking github.com.
Show HN: STE80 Humanizer – de-AI text by writing 80% of the way to ASD-STE100
39 points and 11 comments on Hacker News, linking github.com.
Show HN: Flide: Make Jev-like decision models up to 9.5x faster
99 points and 24 comments on Hacker News, linking github.com.
Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU
Llama-modchip runs DeepSeek-V4-Flash and 30B MoEs on 64GB RAM without GPUs, achieving up to 111 tokens per second. ·
Infrastructure
5 storiesData centers, power, chips and what tokens cost
Data-center and energy stories per day
Table
| Wed 04 | 40 |
| Thu 05 | 36 |
| Fri 06 | 25 |
| Sat 07 | 1 |
| Sun 08 | 2 |
| Mon 09 | 35 |
| Tue 10 | 0 |
Stories from DCD, Data Center Knowledge, EIA and Hacker News, by publication day.
Output price, $ per million tokens (newest models)
Table
| claude-fable-5.2 | 40 |
| claude-fable-5.2:batch | 20 |
| muse-ember | 0.5 |
| lfm-3-1.7b | 0.2 |
| gemini-omni-1.2-flash | 9 |
| median, all paid models | 2 |
438 paid models listed; median of the 32 added in 30 days: $2.50. Rankings data by OpenRouter, CC BY 4.0.
The same H100 costs $2.69 or $12.29 an hour depending where you rent it
164 points and 84 comments on Hacker News, linking github.com.
Google Adds TPU 8tx, 8i Max for Enterprise AI
Google Cloud adds TPU 8tx and 8i Max chips for secure AI workloads and high-capacity inference with large context windows. ·
Nvidia Probes Huang Ties as AI Industry Weighs the Fallout
From Data Center Knowledge.
The same H100 costs $2.69 or $12.29 an hour depending where you rent it
164 points and 84 comments on Hacker News, linking github.com.
Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU
Llama-modchip runs DeepSeek-V4-Flash and 30B MoEs on 64GB RAM without GPUs, achieving up to 111 tokens per second. ·
Google Adds TPU 8tx, 8i Max for Enterprise AI
Google Cloud adds TPU 8tx and 8i Max chips for secure AI workloads and high-capacity inference with large context windows. ·
Nvidia Probes Huang Ties as AI Industry Weighs the Fallout
From Data Center Knowledge.