frontpage.
newsnewestaskshowjobs

Open Source @Github

fp.

Show HN: Distilling DeepSeek into GPT-OSS doesn't transfer censorship. Try it

https://www.ctgt.ai/research/distillation-censorship-transfer
20•cgorlla•1h ago•10 comments

Show HN: Optimize and serve models with Fable quality at half the cost

https://github.com/experientiallabs/world-model-optimizer
50•SilenN•3d ago•22 comments

Show HN: Supapool – a Supabase per coding agent in ~400 ms

https://supapool.io/
12•kevo1ution•1d ago•0 comments

Show HN: Noisegate – a differential-privacy gateway for untrusted AI agents

https://github.com/yashmahajan10/llm-differential-privacy-gateway
8•yashmahajan10•2h ago•0 comments

Show HN: Claude-account – switch Claude Code accounts without logging in again

https://github.com/hamzarehmandeveloper/claude-account
37•hamza_rehman•5h ago•22 comments

Show HN: I made a game where you build a CPU from logic gates

https://select.supply/game/chipbuilder
69•laurentiurad•8h ago•60 comments

Show HN: Tally – check a spreadsheet's numbers against their source, in-browser

https://tally.jiegou.ai/
4•smperng•2h ago•0 comments

Show HN: NegativeEV – I built a tool to help people see how bad their bets are

https://negativeev.com/
2•qkwrv•1h ago•1 comments

Show HN: I'm a senior estimator, I open sourced my construction takeoff tool

https://github.com/Kentucky-ai/opentakeoff
3•kentucky-ai•1h ago•0 comments

Show HN: Leia – a CLI to oneshot playlists onto Yoto MYO cards

https://github.com/nsokin/leia
2•nsokin•1h ago•0 comments

Show HN: APIgen – Generate Fusio REST APIs and Angular UIs from TypeSchema

https://apigen.app/
2•k42b3•1h ago•0 comments

Show HN: State of the Feed – TikTok usage analyzed

https://wrapped.vantezzen.io/state-of-the-feed
2•bennett_dev•1h ago•0 comments

Show HN: Otzar – Local hybrid search over your own files

https://github.com/danielfleischer/otzar
2•yruthewaythatur•1h ago•0 comments

Show HN: Reply to Your Comments with an API

https://mallary.ai/
4•samteeeee•1h ago•2 comments

Show HN: Doodlea.day – A daily step-by-step marker drawing lesson

https://doodlea.day/tutorials/cartoon-soda-can-with-fizz.html
2•mybbor•1h ago•1 comments

Show HN: Ski – Voice Coding for Claude Code, Codex and More – On-Device – Free

https://heyski.io/
9•jomon003•1h ago•5 comments

Show HN: An online Live face swap app, no GPU need

https://livefaceswap.ai/
3•mixfox•2h ago•0 comments

Show HN: Open-source engine running Gemma 4 26B in 2 GB RAM on any M-series Mac

https://github.com/drumih/turbo-fieldfare
875•gitpusher42•1d ago•315 comments

Show HN: ZeroShot: Agent session monitoring to make your team go faster

https://tryzeroshot.com
14•chaos_emergent•2h ago•3 comments

Show HN: secure FFmpeg

2•finbot•3h ago•4 comments

Show HN: I was able to run tiberian sun on cncnet on an M4 MacBook Pro

https://github.com/piyiotisk/cncnet-ts-client-package/blob/master/docs/running-tiberian-sun-on-ap...
2•piyiotisk•3h ago•0 comments

Show HN: Hacker Fables – A satirical cyberpunk novel you can read as a man page

https://sebastiancarlos.github.io/hacker-fables/
4•httbs•6h ago•0 comments

Show HN: CheapFoodMap – A map of good meals under $10

https://cheapfoodmap.com/
269•jaep1•1d ago•239 comments

Show HN: XQL, a query language for screening market data

https://xcreener.com/xql-playground
3•guzzul•6h ago•2 comments

Show HN: Qwen Scribe – local transcription and dictation for Apple Silicon

https://github.com/VladUZH/qwen-scribe
92•sidclaw•1d ago•20 comments

Show HN: OpenMetaHarness – long-horizon execution over multiple context sessions

https://github.com/openmetaharness/openmetaharness
3•IzhaarTejani•4h ago•0 comments

Show HN: A local merge queue for parallel Claude Code agents

https://github.com/funador/claude-code-merge-queue
40•funador•19h ago•22 comments

Show HN: Agentic data transformation & analytic for corporate finance

https://trysecondstate.com/
3•ELAmrani•4h ago•3 comments

Show HN: AgentCouch – let your agents chat with other agents

https://agentcouch.dev
6•spstoyanov•4h ago•1 comments

Show HN: Tilery-VM – run Nvidia cuTile GPU kernels on a CPU, no GPU required

https://github.com/drbh/tilery-vm
4•justdrbh•4h ago•0 comments
Open in hackernews

Show HN: Distilling DeepSeek into GPT-OSS doesn't transfer censorship. Try it

https://www.ctgt.ai/research/distillation-censorship-transfer
20•cgorlla•1h ago
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

ljlolel•33m ago
so interesting!!
data-ottawa•26m ago
FYI the scrolling on iPad with trackpad is broken. A full swipe on the trackpad is about 1 inch of screen movement.
cgorlla•19m ago
Thanks for letting us know. Is this is on the research post?
Alifatisk•25m ago
I’m thinking this makes fullt sense because distillation is only additive, not subtractive. So it does not remove knowledge (if we can define censorship as removal of knowledge).
cgorlla•14m ago
Consider that LLMs are trained on the corpus of the internet, and (simplifying) consequently give the average answer of the internet. If the desired answer of the censorer is contradictory to this, then it requires additional training data to get the model to act a certain way.
andy99•19m ago
So, there is no subliminal learning in this situation, under what conditions would we expect it. I find a transfer attack to be a bit far fetched but it’s definitely interesting.

If we trained from random initialisations on DeepSeek output (that didn’t explicitly contain the political questions) we would expect transfer? And if we fine tuned a model pretrained elsewhere on Deepseek output?

What is the line?

cgorlla•7m ago
It's most likely to occur when distilling a Chinese model from a Chinese base. We plan to do compliance geometry analysis in the future to see what is structurally changing in the model when distillation causes it to start refusing or whitewashing.
dluan•19m ago
It'd be interesting to use this technique to create a running tally across all models of which models are censored on what topics
seri4l•11m ago
Deepseek is, with difference, the most "Western" of Chinese models, so it's a bit perplexing that it was chosen to test this hypothesis.

I didn't run any benchmarks but I played around a little, and after getting around the API-level filter Deepseek V4's answers about "China-sensitive content" aren't any different from what I get from Claude and ChatGPT.

martini333•2m ago
Hijacking scroll behaviour in 2026 is wild.