frontpage.
newsnewestaskshowjobs

Open Source @Github

fp.

Open in hackernews

Show HN: Cactus Hybrid: We taught Gemma 4 to know when it's wrong

https://github.com/cactus-compute/cactus-hybrid
1•HenryNdubuaku•46m ago
Hey HN, Henry & Roman here from Cactus.

A small, on-device model is fast and private, but sometimes wrong, but frontier models are getting expensive pretty fast. So, we post-trained Gemma 4 E2B post-trained to know when it's wrong. Every response comes with a confidence score between 0 and 1. Developers can accept the on-device when it's high, hand off to a bigger cloud model when it's low. By routing only 15-35% of queries to Gemini 3.1 Flash-Lite, Gemma-4-E2B matches Gemini 3.1 Flash-Lite on most benchmarks.

- ChartQA: 15-20%

- LibriSpeech: 25-30%

- MMBench, GigaSpeech, MMAU: 30-35%

- MMLU-Pro: 45-55%

We were always frustrated by the routing signals hybrid apps rely on: asking the model to rate itself in text (unreliable, and you're parsing prose), or token entropy heuristics (barely better than a coin flip in our tests). So we did mechanistic studies on small models, Gemma 4 particularly, and found the hidden state for different layers carry meaningful self-awareness signal for various situations.

SO we extended the model with a 68k params probe layer (LayerNorm, low-rank projection, attention pooling, small MLP head) reads one intermediate layer during decoding and predicts p(wrong); confidence = 1 - p(wrong), returned as structured data, never parsed out of the answer text.

Across 12 hold-out benchmarks spanning text, vision and audio, the probe averages 0.814 AUROC vs 0.549 for token entropy. The result that convinced us this is real: the probe was trained on zero audio data, yet scores 0.79-0.88 AUROC on four audio benchmarks where entropy is near-random or worse (0.32-0.52). It's reading a modality-independent correctness signal from the hidden state, not memorizing patterns from its training data.

We published all weights on HuggingFace and provide copy-pase codes to run it on Transformers, MLX, Llama.cpp or Cactus. With Ollama, vLLM, SGLang etc in the works. For llama.cpp we ship a patch series you compile in once (upstreaming is planned). The code is MIT licensed; Gemma model use remains subject to the Gemma terms.

GitHub: https://github.com/cactus-compute/cactus-hybrid

Weights: https://huggingface.co/collections/Cactus-Compute/cactus-hyb...

Some caveats:

- The probe scores single-sequence decoding only, up to the first 1024 generated tokens.

- Handoff works best when routing per task in a multi-step process, not per step.

- Hierarchical routing is still in the works: try on-device, then DeepSeek v4 Flash, before Fable/GPT5.5/Gemini/Muse/Grok.

- The technique is boutique for each model, we will share each weights as they roll out.

These issues are currently being tackled at Cactus and updated weights will be shipped directly into the HuggingFace collection and GitHub repository straight up. Please let us know your thoughts, it helps us find ways to improve the design progressively.

Thanks a million!

An OpenAI test model escaped and broke into a real company's servers

https://www.cnn.com/2026/07/22/tech/openai-hugging-face-ai-cybersecurity
1•stevenjgarner•41s ago•0 comments

Rabbitty – a native Mac terminal for running AI coding agents in parallel

https://github.com/mauscoelho/rabbitty-app/releases
1•mauscoelho•46s ago•1 comments

Unity 7 Announcement

https://unity.com/releases/unity-7
2•suprfnk•1m ago•0 comments

SimpleX Public Names – A Name Nobody Can Take from You

https://simplex.chat/blog/20260722-simplex-public-names.html
1•Sami_Lehtinen•1m ago•0 comments

Wizardry, the Landmark RPG That Inspired Dragon Quest and Final Fantasy

https://www.timeextension.com/features/the-making-of-wizardry-the-landmark-rpg-that-inspired-drag...
1•homarp•1m ago•0 comments

Digital Surveillance Reshapes Fishery Enforcement in Indonesia

https://spectrum.ieee.org/fishery-satellite-surveillance
2•Jimmc414•2m ago•0 comments

SuperPass: Fast-Tracking Blocking Threads to Mitigate Priority Inversion

https://arxiv.org/abs/2607.18097
3•Jimmc414•3m ago•0 comments

Japan cities hit 40 C, officially marking first "kokushobi"

https://english.kyodonews.net/articles/-/80372
2•geox•4m ago•0 comments

State Reverses Course, Finds Cuyahoga Jail Staff Failed to Start CPR

https://www.themarshallproject.org/2026/07/14/cuyahoga-county-jail-deaths-ohio
2•Jimmc414•4m ago•0 comments

Code mode is all you need: Why agents writing code > calling tools

https://www.browserbase.com/blog/code-mode-is-all-you-need
3•Kylejeong21•5m ago•1 comments

Agent Native Interface for SPI Flashing

https://github.com/jackulau/ratchet
2•JackLau•6m ago•0 comments

Amazon cuts some jobs in its artificial general intelligence unit

https://www.cnbc.com/2026/07/22/amazon-lays-off-some-employees-in-its-agi-unit.html
2•petethomas•6m ago•0 comments

LAN Orangutan – Lightweight network scanner with persistence and Tailscale

https://github.com/291-Group/LAN-Orangutan
2•291Group•8m ago•0 comments

New quantum gravity theory links entropy, dark energy, and life

https://www.sciencedaily.com/releases/2026/07/260719035947.htm
1•aard•11m ago•0 comments

If you pay a hacker's ransom, chances are that they'll come back for more

https://techcrunch.com/2026/07/22/if-you-pay-a-hackers-ransom-chances-are-that-theyll-come-back-f...
1•speckx•11m ago•0 comments

Reddit stock sinks as company weighs blocking Google's AI access to its content

https://qz.com/reddit-stock-google-ai-content-deal-072226
4•01-_-•15m ago•0 comments

Opbox – multiplayer sync for text files on disk

https://www.opbox.dev/
1•sensodine•15m ago•0 comments

AI tech companies have 'hidden debt' worth around $1.65T

https://www.tomshardware.com/tech-industry/big-tech/ai-tech-companies-have-hidden-debt-worth-arou...
6•01-_-•16m ago•0 comments

When your vehicle outlives its cloud: What happens next?

https://arstechnica.com/cars/2026/07/when-your-vehicle-outlives-its-cloud-what-happens-next/
2•apparent•17m ago•0 comments

Jacobian Conjecture for Baby

https://muchmirul.github.io/jacobian-conjecture/
2•porphyra•18m ago•0 comments

Using AI makes people less likely to admit they don't know something

https://www.theregister.com/ai-and-ml/2026/07/19/using-ai-makes-people-less-likely-to-admit-they-...
2•smartmic•20m ago•0 comments

Navigating the volatile silicon market: updates on memory and storage pricing

https://frame.work/blog/updates-on-memory-pricing-and-navigating-the-volatile-memory-market
2•davidjade•21m ago•0 comments

What to Expect from China's New Five-Year Plan

https://institute.global/insights/geopolitics-and-security/what-to-expect-from-chinas-new-five-ye...
1•toomuchtodo•21m ago•1 comments

RefluXFS: LPE in the Linux kernel via XFS reflink race (CVE-2026-64600)

https://www.openwall.com/lists/oss-security/2026/07/22/14
2•ilreb•23m ago•0 comments

Amazon once again locking me out of my own order data

https://bsky.app/profile/jon.moeller.io/post/3mrawmjzsmc2u
3•jmole•24m ago•2 comments

It's time to wake up: An Emacs alarm clock

https://lars.ingebrigtsen.no/2026/07/22/its-time-to-wake-up-an-emacs-alarm-clock/
1•meetpateltech•24m ago•0 comments

AI-maestro: Conduct a roster of AI coding agents against a work board

https://github.com/my-chiefmind/ai-maestro
1•mychiefmind•25m ago•0 comments

Comhad: A TUI replacement for Cyberduck, written in Rust

https://github.com/Eoin-McMahon/comhad
1•emcmahon478•25m ago•0 comments

The Healthcare Industry Is Coming for Your Face

https://www.karenschoellkopf.com/writing/biometric-data-healthcare
2•kgunette•25m ago•0 comments

Low-volume wearable interface for voice computing

https://twitter.com/augmentaltech/status/2079962940896338039
2•tomasero•27m ago•0 comments