GigaToken: ~1000x faster Language model tokenization

<a href="https:&#x2F;&#x2F;book-prize-index.vercel.app&#x2F;" rel="nofollow">https:&#x2F;&#x2F;book-prize-index.vercel.app&#x2F;</a>

Quality non-fiction books are the antithesis of AI slop

Terrence Tao's ChatGPT Conversation about the Jacobian Conjecture Counterexample

Malleable Computing, Emacs, and You

Safari Technology Preview 248 Released

Are AI Labs Pelicanmaxxing?

<a href="https:&#x2F;&#x2F;xcancel.com&#x2F;na_announce&#x2F;status&#x2F;2079952538040672302" rel="nofollow">https:&#x2F;&#x2F;xcancel.com&#x2F;na_announce&#x2F;status&#x2F;2079952538040672302</a>

John C. Dvorak has died

Hey HN, Henry &amp; 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&#x27;s wrong. Every response comes with a confidence score between 0 and 1. Developers can accept the on-device when it&#x27;s high, hand off to a bigger cloud model when it&#x27;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&#x27;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&#x27;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: <a href="https:&#x2F;&#x2F;github.com&#x2F;cactus-compute&#x2F;cactus-hybrid" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;cactus-compute&#x2F;cactus-hybrid</a>Weights: <a href="https:&#x2F;&#x2F;huggingface.co&#x2F;collections&#x2F;Cactus-Compute&#x2F;cactus-hybrid-6a60da4551074db058e8bb64" rel="nofollow">https:&#x2F;&#x2F;huggingface.co&#x2F;collections&#x2F;Cactus-Compute&#x2F;cactus-hyb...</a>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&#x2F;GPT5.5&#x2F;Gemini&#x2F;Muse&#x2F;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!

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

Nobody knows what a used GPU cluster is worth

Hi HN, We’re Adeel and Umair, co-founders of Unlayer (<a href="https:&#x2F;&#x2F;unlayer.com&#x2F;">https:&#x2F;&#x2F;unlayer.com&#x2F;</a>). We let you add content creation to your applications without having to build an entire editor, renderer, template, and export stack yourself. Unlayer lets you create emails, web pages, and documents inside your app, in three different ways: in code, visually, or with AI.Here’s a demo: <a href="https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=0HsDtNkdMpM" rel="nofollow">https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=0HsDtNkdMpM</a>.We started with an embeddable email editor because a lot of products eventually need one: CRMs, marketing tools, customer engagement platforms, marketplaces, internal tools, and vertical SaaS apps all run into this at some point. At first, it sounds like a small feature: &quot;just&quot; add a drag and drop editor. In practice, it turns into a big pain. You end up dealing with email rendering, Outlook quirks, responsive layouts, templates, merge tags, image uploads, exports, permissions, localization, versioning, and a long tail of edge cases that have nothing to do with your core productOver time, we saw the same problem beyond email. Apps also need landing pages, invoices, proposals, reports, contracts, and PDFs. Some of this content is best created visually by end users. Some of it is better generated in code by developers. Increasingly, some of it is also generated by AI agents. Many teams eventually need all three workflows. That is the direction we have been working toward with Unlayer.There are three parts we are showing today:(1) Unlayer Elements. This is our open-source React component library for creating emails, pages, and documents in code (repo: <a href="https:&#x2F;&#x2F;github.com&#x2F;unlayer&#x2F;elements" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;unlayer&#x2F;elements</a>, more at <a href="https:&#x2F;&#x2F;unlayer.com&#x2F;elements">https:&#x2F;&#x2F;unlayer.com&#x2F;elements</a>). Instead of hand-writing raw HTML templates, developers can compose content using React components, reuse sections like headers, footers, CTAs, invoice rows, and branded blocks, keep templates in Git, and render them into production output.One newer use case we are seeing is AI-assisted content creation. If an AI agent is asked to create an email, invoice, report, or landing page, the output is usually raw HTML or markdown that becomes hard to maintain. With Elements, the agent can generate structured React components instead. A developer can review the result, refactor it, keep it in Git, and still pass the design into a visual builder later if someone needs to edit it.(2) Visual Builder. This is the drag and drop editor (repo: <a href="https:&#x2F;&#x2F;github.com&#x2F;unlayer&#x2F;react-email-editor" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;unlayer&#x2F;react-email-editor</a>, more at <a href="https:&#x2F;&#x2F;unlayer.com&#x2F;email-builder">https:&#x2F;&#x2F;unlayer.com&#x2F;email-builder</a>) that can be embedded inside an app so non-technical users can create or edit content. In the demo, we show the email builder and the AI assistant inside the builder. The goal is not to replace the developer workflow, but to connect it with a visual workflow when marketers, admins, customers, or internal teams need to make changes themselves.(3) Document Builder. This is for structured documents such as proposals, reports, invoices, contracts, and PDFs. We have seen a lot of teams build separate systems for email templates, web pages, and document generation, even though the underlying primitives are similar: layout, content blocks, variables, assets, preview, export, and permissions. More: <a href="https:&#x2F;&#x2F;unlayer.com&#x2F;document-builder">https:&#x2F;&#x2F;unlayer.com&#x2F;document-builder</a>The technical challenge is making these workflows share a common foundation. Developers should be able to build templates in code when that makes sense. End users should be able to edit visually when that makes sense. AI agents should be able to generate structured content instead of unmaintainable blobs. The final output should still be usable by the host application.We make money by selling hosted builder, template, export, and platform features to companies embedding this into their products. Elements is open source. The commercial product is the broader hosted platform around builders, collaboration, storage, exports, and production use cases.We were part of W22, so this is a late Launch HN. At the time, Unlayer was an embeddable email editor, and we did not think we had the right broader story for HN yet. Since then, the product has expanded into a more general content creation layer for products, including emails, pages, documents, APIs, open-source developer projects, and AI-assisted workflows. That felt like a better moment to bring it to HN and ask for feedback.We&#x27;d really appreciate thoughts from HN. This is one of those areas where a lot of people have strong opinions because they’ve been burned by editors, email HTML, document editing, or “simple” content workflows before. We&#x27;d love to hear what resonates, what sounds wrong, and what you think we should be thinking harder about!

Launch HN: Unlayer (YC W22) – Add email and document builders to your app

The startup's Postgres survival guide

How we made our LeRobot video reader up to 15× faster

Businesses with ugly AI menu redesigns

Nvidia DGX Spark as a daily driver

Ghost Cut – or why Cut and Paste is broken everywhere

I&#x27;m the author of a paper my friends and I wrote after we were curious if a MUD, text games originating in the 1970s, could be used to evaluate LLMs. We&#x27;ve spent the last several months on nights and weekends running this experiment and writing the paper on just our personal computers with about $99 in API credits.Our experiment did have an interesting leaderboard but even more surprising was the measurements of each LLM. We scored each on four behavioral dimensions, two of which lean heavily on an LLM classifier. When we removed those two, one of the frontier models fell six positions. When we then checked the classifier against a second judge, the per-model agreement between them ranged from 85% to 22%. The aggregate kappa (0.04 on probe detection) indicated the instrument was noisy without saying which models the noise was hitting. The most affected model shared a model family with the classifier. This isn&#x27;t proof of bias, just one observation we recorded.We realize LLM judges can be unreliable, and while it wasn&#x27;t our original intent to test this, it ended up being the most interesting finding. The divergence between the two judges is the finding we think generalizes to other judge-based benchmarks.We emphasize this is just a proof of concept and not a validated benchmark. We prepared a thorough limitations section in the paper, including just 50 runs per model, overlapping CIs among the top models, no human raters, a tiny environment, etc.Everything we did is publicly available, the paper and data are CC BY 4.0, while the code is MIT. The paper, transcripts, code, and complete API billing export can be found at <a href="https:&#x2F;&#x2F;doi.org&#x2F;10.5281&#x2F;zenodo.21386663" rel="nofollow">https:&#x2F;&#x2F;doi.org&#x2F;10.5281&#x2F;zenodo.21386663</a>If you find issues, please let us know, that&#x27;s why we&#x27;re sharing this. We&#x27;re currently designing Phase 2 and want it to be as robust as possible. We&#x27;re looking at human baselines, multiple judges, more objectives, a larger environment, etc.

Can a MUD evaluate LLMs? A $99 proof of concept

Hey,I would like to show you my project, but it&#x27;s difficult because it only works with a registration, so I explain the concept to you.The Idea:
Everyone who registers has to do an onboarding and answer meaningful questions that help you to find a right match. The matching happens by the system. Once you&#x27;re done with the onboarding, you enter the pool. When a match is found, you go into a 1on1 14 question-set with that person. All answers are revealed immediately. At the end both decide if they form a connection or not. Only if both agree, the chat opens. Once a connection is formed, they can play through other prepared question sets, so the conversation doesn&#x27;t get dry. It&#x27;s also possible to create custom question sets to play through with your connected matches.I just got this idea because I really wanted away from all the typical friendship and dating apps that are just a marketplace of user profiles. I feel like that&#x27;s the only hook other apps offer, people come back because they are bored and browse through profiles. On tinder it&#x27;s swiping through pictures and on other apps it&#x27;s a list of profiles where you can filter with your matching criteria. But all that felt just shallow to me. I think people rarely ask the right questions each other, like really mandatory things that matter to form a any kind of relationship with someone. I&#x27;ve been on multiple different social platforms and it always felt mega shallow. Most people can&#x27;t even start a good conversation, it&#x27;s often just stuck at small talk. Maybe it&#x27;s not even peoples fault, that&#x27;s why I tried with these question sets to tackle multiple problems. First, that you don&#x27;t even have to bother with someone if you have entirely different worldviews, which you would normally just find out if you talk long enough with someone. And second, just killing the small talk issue, by giving people a base.So before I started coding this project, I actually did a research if there is anything backed by science to find out, which kind of questions would actually matter to find out, if you can handle someone. There were a lot of studies I&#x27;ve found e.g. &quot;Aron and colleagues (1997)&quot;, &quot;Hall (2019)&quot;, &quot;Byrne (1971)&quot;, &quot;Finkel and colleagues (2012)&quot; and &quot;Hazan and Shaver (1987)&quot;. I took their research work and tried to create questions based on them.The project is running on Next, Postgresql, everything dockerized, running on a small Hetzner VPS.
I used Claude+Codex to develop everything, but I wouldn&#x27;t say it&#x27;s really vibecoded, since I work as a Senior Dev and I used my main tech stack.It&#x27;s an open beta currently and it&#x27;s live for like 14 days - I have ~150 users right now. And honestly, I am really overwhelmed by those numbers, because it&#x27;s the first time that I ever managed to attract this amount of people on a project I&#x27;ve launched.When I started, I launched it as a closed beta, so it was invite only. During that period I got so much good feedback from people, it was incredible. I implemented almost everything my first users told me. Since all the changes, I didn&#x27;t retrieve any further feedback what people wish would be different. It&#x27;s just constant good feedback since then. People contact me from alone and just tell me how much they like these questions and the loop.My marketing strategy is mainly Reddit currently, but I started also new methods. I found some super relevant subreddits that had no rules against advertising. When I wrote my posts on reddit, I didn&#x27;t mention a product name or link. I&#x27;ve only described the idea and if somebody would have interest trying it out. That worked surprisingly well.
Another thing I&#x27;ve done was using a social app that allows anonymous posts at your current location, kinda something like &quot;Nextdoor&quot;.
Posts on this app bring me daily 1-5 new registrations.I wonder what you guys think about it and I am open for any critics!

Show HN: ValuePair – a friendship app that cares about values first

Does creatine make you smarter?

<a href="https:&#x2F;&#x2F;xcancel.com&#x2F;mkratsios47&#x2F;status&#x2F;2079933645888880708" rel="nofollow">https:&#x2F;&#x2F;xcancel.com&#x2F;mkratsios47&#x2F;status&#x2F;2079933645888880708</a>

“We have information that Moonshot distilled Fable for the development of K3”

Full Scale Foldable Wing Extensions

Hi HN - I&#x27;m Venkat, founder of Stayflexi (YC), CMU CS grad and Ex-Oracle Query Engine team (patents in core databases)DeepSQL started as an internal tool to stop our own databases from becoming the bottleneck they were becoming (13,000+ hotels in production). It worked well enough that we&#x27;re releasing it.DeepSQL is an AI agent that operates a database the way a senior DBA and Data Engineer would1. Fixes slow queries (we cutdown DB spend by 4x)
2. Fixes DB bloat (blocks unnecessary schema changes, in vibecoded setup)
2. BI dashboards(we removed spend on tableau, retool and appsmith)
3. Security (We could redact access to sensitive, PII data to our employees) while engaging everyone in the org to interact with our databaseHow it works
1. Deepsql learns your data and relationships from your code base, rules and query logs
2. Deepsql agent has 20 background jobs that continuously monitor the schema changes, bottlenecks, and come up with solutions before the problem arises. (Schema bloat issues are irreversible)
3. Deepsql Brain has CLI and MCP surfaces that can directly work with Claude &#x2F; Codex and Cursor. 
4. Deepsql digest delivers the daily database health report.
5. Your team can connect to deepsql with their favourate surface areas - Web UI, CLI, MCP or Slack.One line install:
curl -fsSL <a href="https:&#x2F;&#x2F;install.deepsql.ai&#x2F;install.sh" rel="nofollow">https:&#x2F;&#x2F;install.deepsql.ai&#x2F;install.sh</a> | bashFor community support:
<a href="https:&#x2F;&#x2F;discord.gg&#x2F;duEJq7AeeG" rel="nofollow">https:&#x2F;&#x2F;discord.gg&#x2F;duEJq7AeeG</a>Or just email me for expert level setup guidance
venkat@deepsql.ai

Show HN: DeepSQL – A self-hostable DBA agent for Postgres and MySQL

Intel Starts Shipping High-NA EUV Silicon

<a href="https:&#x2F;&#x2F;xcancel.com&#x2F;nikitabier&#x2F;status&#x2F;2079787406300266743" rel="nofollow">https:&#x2F;&#x2F;xcancel.com&#x2F;nikitabier&#x2F;status&#x2F;2079787406300266743</a>

Passkeys were invented by engineers with zero understanding of consumer brain

ScreenWall – Turn old phones into synced widgets for your space

Show HN: HN Hall of Fame – browse 3,100 legendary Hacker News links

I built The Daily FM because I wanted a daily summary of the latest AI news sent to my podcast app and then realized it was useful in summarizing other stuff:- Any blogs, websites, or X handles- The top Hacker News stories with comments- Long podcasts (think Lex Fridman)- Changes in state&#x2F;federal legislationIt pulls your sources, writes a script with a frontier-ish model of your choice, runs it through TTS (MAI-2 studio voices), and publishes a standard RSS feed. You can also combine several pods into one feed for one subscribe point.Under the hood it’s using Cloudflare Workers, queues, browser rendering, and the AI Gateway, but the TTS models aren’t very good on Cloudflare so I switched to the MAI-Voice-2 models with OpenRouter.I also created a public library you can subscribe to without ever signing up, or if you sign up, you can combine a bunch into a single feed and use your own sources as well and choose from other voice options if you don’t like “Excited Ethan” at 1.1x speed.Feedback welcome.

11:00:30 PM WEDNESDAY JULY 22, 2026

THE HACKER NEWS