Hacker News

Article URL: https://www.texturehq.com/blog/i-built-this-for-myself-first

Comments URL: https://news.ycombinator.com/item?id=49113668

Points: 1

# Comments: 0

Article URL: https://utilitly.com/blog/how-to-split-and-extract-pdf-pages-securely

Comments URL: https://news.ycombinator.com/item?id=49113660

Points: 1

# Comments: 0

Article URL: https://www.wsbtv.com/news/local/apalachee-school-shooting-father-shooter-faces-own-sentence-up-180-years/ZXCE3UUUFZH6NPGN2JR5LGRSHM/

Comments URL: https://news.ycombinator.com/item?id=49113643

Points: 5

# Comments: 0

Article URL: https://github.blog/changelog/2026-07-30-reference-same-repository-actions-with-self-repository-syntax/

Comments URL: https://news.ycombinator.com/item?id=49113640

Points: 1

# Comments: 0

Article URL: https://www.isgroup.biz/en/cyber-security/llm-based-code-security-review-costs-findings-methodology.html

Comments URL: https://news.ycombinator.com/item?id=49113630

Points: 1

# Comments: 0

Article URL: https://ti.org/antiplanner/?p=24073

Comments URL: https://news.ycombinator.com/item?id=49113611

Points: 1

# Comments: 0

Article URL: https://www.gilesthomas.com/2026/07/why-do-openai-gpt2-weights-beat-mine-2-the-bugfix

Comments URL: https://news.ycombinator.com/item?id=49113601

Points: 1

# Comments: 0

We recently used DeepSeek V4 Flash as a teacher for finance tasks with GPT-OSS-120B. Distillation works well on this problem. At a constrained 8k token budget, our self-distilled 120B scores 83.61% on FinanceReasoning, above Kimi K3 (81.93%) and Inkling (65.13%). We released the 20B open weights. With V4 as the teacher though, we realized it would be timely to measure if the censorship characteristic of it transferred to the distilled version of the base model. tl;dr it didn't, the teacher answered politically sensitive questions 7 SDs differently than expected, but the distilled model's behavior remained the same as its American base. You can try a couple queries yourself with no auth here: http://playground.ctgt.ai/

I will now dive in to the motivation, methodology and detailed results for those interested. The hard part of measuring this phenomena is isolating whether a model is reluctant to talk about sensitive things generally vs. a particular country's sensitive things. So we made 152 matched pairs where one prompt asked about a Chinese concept, and the other asked about a non-Chinese version of that concept. For example, the Great Leap Forward vs. the Holodomor. These were scored 0-100 by four LLM judges (Grok 4.20, Gemini 3.5 Flash, GPT-5 mini, Claude Sonnet 4.6), validated against 96 human scores at r=0.948. OpenRouter blocked some of these so we hosted the weights ourselves.

The teacher's gap on the core political set of pairs was +45.45 points, ~7 standard deviations from chance, and every distilled student was within 1 point of its base. Subliminal learning literature says this is expected when the initializations are not shared between teacher and student, which is true here. The distillation data also did not contain any China-sensitive content. The contribution here was to release the evaluation framework (LineageEval: https://github.com/CTGT-Inc/lineage-eval/) to elevate the discussion around this topic in DC and beyond. We are an interpretability lab working on high risk and regulated applications of AI, so we hear a lot of vagaries aimed at the supposed dangers of distilling Chinese models on American bases. We believe these conversations should be based on open, auditable frameworks and not feelings. We plan to test what happens with a Chinese teacher into a Chinese-lineage base like Qwen next.

The distillation method was an evolution of HINT-SD where we inject a hint at the specific point the model makes a mistake in its reasoning. Then we train on the corrected continuation with reverse KL over the next 100 toks of the rollout. As mentioned above 120B itself was efficacious as a teacher, and we ended up shipping this version. The self-distilled 120B scores 83.61% on FinanceReasoning, above Kimi K3 (81.93%) and Inkling (65.13%). Ours finishes 98.7% of problems in budget; the larger models truncate (90.76% and 71.01%) which score as incorrect. At 100k tokens big models gain (Kimi 89.92%). So for a finance task at a constrained (perhaps more realistic) budget a 120B on one H100 at ~$0.00026/query outpaced models running 62-160x more per query.

We put out the 20B finance model as open weights (64.71% to 74.79% at 8k on FinanceReasoning, 23% lower cost/query, runs on one 80GB GPU), the 120B in a playground with teacher and students side by side (a few queries, no auth), and LineageEval with all prompts, controls, rubric, and code.

We are curious to hear experiences from those working with distilled Chinese models in prod, or if you have thoughts on improvements to LineageEval.

https://huggingface.co/ctgt-inc/gpt-oss-20b-finance

https://playground.ctgt.ai/

https://github.com/CTGT-Inc/lineage-eval/

https://www.ctgt.ai/research/distillation-censorship-transfe...

Comments URL: https://news.ycombinator.com/item?id=49113599

Points: 3

# Comments: 0

Hey HN, I built ACS Peptide Testing Labs, an ISO/IEC 17025 accredited analytical lab in Florida that lets researchers and compounding pharmacies mail in peptide samples and get a full Certificate of Analysis back within 9–11 business days.

The problem was simple: most peptide suppliers provide their own COAs, which is a obvious conflict of interest. Researchers needed somewhere independent to send samples.

You submit through our web portal, ship overnight, and we run HPLC purity analysis, ESI mass spectrometry for identity confirmation, ICP-MS for heavy metals, and GC-MS for residual solvents. Results come back as a detailed COA via email.

One interesting constraint: we're DEA registered, which lets us handle controlled substance analysis that most contract labs won't touch.

We're still expanding our custom protocol offerings and would genuinely love feedback from anyone working in peptide research or compounding.

Comments URL: https://news.ycombinator.com/item?id=49113597

Points: 2

# Comments: 0

Several weeks ago, I wanted to draw a cowboy hat for a handmade birthday card, so I used ChatGPT to create a drawing tutorial for me. It was impressive how well the image generation was able to break down the end result into a step-by-step lesson, and this led to the idea of creating a daily AI-generated drawing lesson website.

Doodlea.day is that. A fresh drawing lesson published most days.

I've had some challenges keeping the lessons cumulative. For example, in the first few steps, a character's eyes might be looking to the right, but then in the last three steps the eyes swap to looking to the left. I've been tweaking the prompt and guidance to the point where most lessons are one-shot these days. All of the steps are generated together in a single contact sheet rather than one image at a time. Each panel has a controlled job, like silhouette, contour, features, detail, and finish, and there is a validation process that identifies and rejects broken lesson panels.

The project is at a place where I am excited to share it. Taking a little time every day with the family to draw has been a great experience, and I hope others can find the same joy.

Comments URL: https://news.ycombinator.com/item?id=49113588

Points: 1

# Comments: 0

SKI is a on-device voice coding application, which can be used with any agents that supports skill, such as Claude Code, Codex or Hermes.

It transcribes your voice (which you can optionally review and edit) and send it to the connected agent. The agent then completes the task, and uses the skill to talk about the updates of the project or a summary or a query using voice.

This runs completely on-device. Free. No subscription. Available on both Mac (notch and pill) and Windows (pill widget)

It can also be sent to meetings with the intelligence of the connected project to participate actively in the meeting. This is a paid feature (as it runs on the cloud), powered by agentcall and is optional.

All on device functions are free with no limits on usage.

Comments URL: https://news.ycombinator.com/item?id=49113559

Points: 7

# Comments: 3

Article URL: https://pulsefeed.dev/correction

Comments URL: https://news.ycombinator.com/item?id=49113551

Points: 1

# Comments: 0

Article URL: https://github.com/yashmahajan10/llm-differential-privacy-gateway

Comments URL: https://news.ycombinator.com/item?id=49113543

Points: 2

# Comments: 0

Article URL: https://observer.co.uk/news/international/article/mauritania-divorce-party

Comments URL: https://news.ycombinator.com/item?id=49113528

Points: 1

# Comments: 0

Article URL: https://hannahilea.com/blog/birduino/

Comments URL: https://news.ycombinator.com/item?id=49113523

Points: 1

# Comments: 0

Article URL: https://uni-tuebingen.de/en/university/news-and-publications/press-releases/press-releases/article/two-bird-figurines-found-in-world-heritage-site-of-hohle-fels-are-declared-find-of-the-year/

Comments URL: https://news.ycombinator.com/item?id=49113477

Points: 2

# Comments: 0

Article URL: https://www.stevestewartwilliams.com/p/peering-at-the-sun-through-the-earth

Comments URL: https://news.ycombinator.com/item?id=49112100

Points: 1

# Comments: 0

Article URL: https://www.ft.com/content/280336bf-dbed-405f-b38e-5af644a21549

Comments URL: https://news.ycombinator.com/item?id=49112092

Points: 1

# Comments: 0

Article URL: https://www.aljazeera.com/sports/2026/7/30/uefa-members-agree-fifa-boycott-if-world-cup-plans-pursued

Comments URL: https://news.ycombinator.com/item?id=49112089

Points: 1

# Comments: 0

Article URL: https://www.cnbc.com/2026/07/30/amazon-zoox-robotaxi-rides-las-vegas.html

Comments URL: https://news.ycombinator.com/item?id=49112083

Points: 1

# Comments: 0