Can Gemma and Qwen models catch hallucinations by looking at their own logprobs?
Hi! I'm really obsessed with LLM hallucinations for the last 6 days 😭 I started by designing system prompts to attack hallucinations but failed, obviously. Now…
Full feed
Feed 28,055 posts
Every story across every category, newest first. Each card links to the original publisher; daily-brief posts open as editorial pages.
RT khipu.ai: Behind a lot of great AI projects, there is a good @JeffDean story. Here is mine 😉 Back in 2019, even before Khipu AI officially exist…
RT khipu.aiBehind a lot of great AI projects, there is a good @JeffDean story. Here is mine 😉Back in 2019, even before Khipu AI officially existed,…
On-Policy Self-Distillation without Any Supervision
On-policy (Self-)Distillation (OPD / OPSD) has shown strong potential for post-training large language models (LLMs). However, existing methods still rely heavily on external supervision, including ground-truth…
v0.32.0
What's Changed Make it easier to debug nested tensors. by @comfyanonymous in #15383 Update workflow templates to v0.11.37 by @comfyui-wiki in #15415 Minimum officially supported pytorch…
WorldClaw Agentic 3D open-world generation at scale
Article URL: https://tencent-hunyuan.github.io/Hunyuan3D-WorldClaw/ Comments URL: https://news.ycombinator.com/item?id=49265051 Points: 19 # Comments: 5
The least verifiable part of AI R&D (and the thing which may bottleneck the automated researchers) is making calls on large experiments and big traini…
The least verifiable part of AI R&D (and the thing which may bottleneck the automated researchers) is making calls on large experiments and big training runs.Interestingly,…
pg_clickhouse v0.10: Subquery pushdown and 1000x faster TPC-H queries
Article URL: https://clickhouse.com/blog/pg_clickhouse-whats-new-july-2026 Comments URL: https://news.ycombinator.com/item?id=49265031 Points: 5 # Comments: 0
Accel closes oversubscribed $550M India fund within weeks, 19 months after its last
The U.S. VC firm still has more than 55% of its previous $650 million India fund available for deployment.
Accelerate cyber defense with OpenAI and AWS: Daybreak Red & Daybreak Blue now available to eligible customers on Amazon Bedrock
Daybreak Red and Daybreak Blue from OpenAI, specialized cyber defense models from OpenAI, are now available on Amazon Bedrock to eligible customers. Both models run with…
We quantized DeepSeek V4 0731 and benchmarked it against popular quants on 8× RTX 5090
We converted the model from the original safetensors and found two issues. The first one made our quantization fail several times, the second one does not…
Suzanne: AI tool for designing and manufacturing physical products
Article URL: https://www.suzanne3d.com/ Comments URL: https://news.ycombinator.com/item?id=49264755 Points: 4 # Comments: 0
RT Ryan Greenblatt: My median for full automation of AI R&D is around late 2030/early 2031. But my "modal"/best guess prediction for this milestone wo…
RT Ryan GreenblattMy median for full automation of AI R&D is around late 2030/early 2031. But my "modal"/best guess prediction for this milestone would be significantly…
Emergent Introspective Awareness in Large Language Models
Article URL: https://arxiv.org/abs/2601.01828 Comments URL: https://news.ycombinator.com/item?id=49264583 Points: 4 # Comments: 1
Decoupled Descent: Enforcing Exact Train-Test Error Tracking Via AMP Onsager Corrections [R]
Link: https://arxiv.org/pdf/2604.27883 Hi, Most of use are familiar with the headache of training a neural network using gradient descent where the training error may go to…
🔬The BioAI Phase Shift – Matthew McPartlon & Neil Patil, Chai Discovery
Pharma is suddenly paying for Bio × AI tools, and Chai is leading the pack with four deals closed this summer. Cofounder Matt McPartlon and Product…
The 19th-Century Family Fortunes Funding Degrowth
Article URL: https://www.effort.news/p4ne Comments URL: https://news.ycombinator.com/item?id=49264352 Points: 3 # Comments: 1
Ryan believes we could get 3-6 years of AI progress at the current pace, but in a single year, once we automated AI R&D. So a jump as big as GPT 3.5 t…
Ryan believes we could get 3-6 years of AI progress at the current pace, but in a single year, once we automated AI R&D.So a jump…
datasette-upload-dbs 0.5a0
Release: datasette-upload-dbs 0.5a0 This plugin has been around for a while - it lets users upload a brand new SQLite database to a hosted Datasette instance,…
366 t/s Qwen3.6 27B NVFP4 on v100s
These are single stream numbers Following on from my previous post about v100s (here) and inspired by this comment (here) I decided to work on kernels…
Claude Opus 5 Just Beat My Text-Based Adventure Game Benchmark
Cross-posted from my Substack. Basically, I created a text-based adventure game benchmark in April, and this morning my agent harness using Claude Opus 5 solved it…
Copilot memory and Ollama in GitHub Copilot for JetBrains
This update brings persistent memory, local model access, and more enterprise controls to GitHub Copilot for JetBrains. It also improves everyday chat workflows and resolves reliability…
b10380: chat : fix muse-glimmer detection of tool calls after EOM (#26879)
chat : fix muse-glimmer swallowing a trailing tool call into content Muse Glimmer routinely answers the user and calls a tool in a single generation. The…
Local Benchmark : Muse Glimmer 30B vs Qwen 3.6 27B vs Gemma4 31B (and many other models and finetunes)
Needs a lot of requests compared to Qwen (almost twice) and Gemma (almost x3). Final score is fine, even though it is "not a coding model"…
Cultivar: A Contrastive and Locale-Oriented Translation Benchmark for Investigating Contamination and Localisation Robustness
Multilingual translation benchmarks are typically sourced in English and translated into other languages, treating language pairs as the unit of evaluation---a design that is prone to…