Hebbia Shows What "Can't Miss a Detail" Actually Means for Frontier AI in Finance
π‘ INSIGHT
Hebbia Shows What "Can't Miss a Detail" Actually Means for Frontier AI in Finance
Anthropic's latest case study pulls back the curtain on Hebbia, a financial diligence platform where missing a single buried clause in a 500-page credit agreement can cost millions. Hebbia runs Claude across full document sets β not summarized chunks β to surface every relevant risk factor, covenant, and obligation. The takeaway for anyone building in regulated industries: the bar isn't "pretty good extraction," it's "zero misses," and that's where long-context frontier models finally start earning their keep over retrieval-augmented patchwork. Read more β
Samsung Tells Health App Users: Let Us Train AI on Your Data or We Delete It
Samsung is pushing a consent screen in its Health app that gives users two choices: agree to let Samsung use your health data for AI training, or have your data deleted. (225 likes | 62 RTs) There's no "keep my data but don't train on it" option β a textbook coercive consent pattern that privacy regulators in the EU and South Korea are almost certainly going to scrutinize. If you're building any app that touches health or biometric data, audit your own consent flows now β because whatever enforcement action Samsung draws will set the template for everyone else. Read more β
π§ LAUNCH
llama.cpp Cracks Open Hunyuan V3 with Multi-Token Speculative Decoding
llama.cpp b9993 lands Hy3 architecture support, meaning Tencent's massive Hunyuan V3 MoE model now runs locally with multi-token prediction speculative decoding. In the same release run: b9990 adds Minimax2 eagle3 spec decoding, and b9992 refactors CUDA MMQ kernels for Blackwell GPUs. The pace here is staggering β architectures that were cloud-only months ago are becoming weekend experiments on local hardware, and speculative decoding is the engine making it practical. Read more β
Qwythos 9B Hits 45K Downloads as Local Multimodal Converges on the Sweet Spot. Qwythos-9B-v2 is trending on HuggingFace with 45.2K downloads β a multimodal image-text-to-text model in GGUF format that confirms what the download numbers have been saying for months: the practical sweet spot for local multimodal inference is the 7β9B parameter range. Big enough to be useful, small enough to run on a single GPU. (102 likes | 45.2K downloads) Read more β
π§ TOOL
MCP TypeScript SDK Splits Its Core for the 2.0 Era. The MCP TypeScript SDK v2.0.0-beta.4 extracts spec schemas, OAuth schemas, and protocol constants into a shared @modelcontextprotocol/core package. If you're building MCP servers, your imports will break before the 2.0 stable release β start migrating now while it's still beta. Read more β
Vercel AI SDK Ships Four Patches in Three Days, Fixes a Streaming Memory Leak. Vercel AI SDK v7.0.23β7.0.26 landed fast: v7.0.26 groups orphaned tool calls under parent spans for cleaner tracing, v7.0.25 fixes a streaming transcription memory leak where cancelled streams didn't abort pending connections, and v7.0.23 wraps embedMany in tracing context. If you use streaming transcription, update immediately β that memory leak is the kind that only shows up in production under load. Read more β
Next.js 16.3 Preview Graduates Request Insights, Fixes Turbopack CSS Ordering. Next.js v16.3.0-preview.6 promotes Request Insights from canary to preview and ships a Turbopack fix that orders CSS modules by chunk-group co-occurrence β resolving the styling inconsistencies that have been blocking CSS Modules adoption in Turbopack builds. If CSS ordering was your blocker, test this preview. Read more β
Jacquard: A Language Designed from Scratch for AI-Written, Human-Reviewed Code. Jacquard is a new programming language built explicitly for the AI-writes-human-reviews workflow. The bet: existing languages carry too much legacy syntax that makes AI-generated code hard for humans to audit quickly. Whether purpose-built languages for AI generation gain traction or whether better tooling on existing languages wins is an open question β but the design rationale is worth reading either way. (37 likes | 13 RTs) Read more β
π¬ RESEARCH
Academic Study Examines How Microsoft Deployed Claude Code Alongside Copilot CLI. A new paper documents Microsoft's early-2026 internal rollout of Claude Code alongside GitHub Copilot CLI β rare empirical data on how a major tech company evaluates competing AI coding tools side by side. If you're making enterprise AI tool adoption decisions, this is one of the few studies with real deployment data rather than benchmark comparisons. (10 likes | 2 RTs) Read more β
MIT Finds a Way to Audit AI Training Data for CSAM Without Generating Any. MIT researchers developed a technique to detect whether AI models were trained on child sexual abuse material β critically, without generating any harmful content in the process. This sidesteps the ethical trap where verification requires reproduction, and it's a significant advance in AI safety auditing methodology. (11 likes | 5 RTs) Read more β
π TECHNIQUE
Someone Built a Working Neural Network in Pure SQL. A fully functional neural network implemented entirely in SQL β forward pass, backprop, the works. It's more art than engineering recommendation, but it's a genuine demonstration of SQL's computational completeness and a reminder that "can this compute?" and "should this compute?" are very different questions. (46 likes | 12 RTs) Read more β
The Practical Guide to Running RL Training at 4-Bit NVFP4 Precision. If you're pushing reinforcement learning onto consumer GPUs, this deep dive on NVFP4 quantization for RL training documents exactly where the stabilityβperformance cliff edges are. The key finding: 4-bit works for RL, but you need specific techniques to keep gradient signals from collapsing β and the paper walks through each one. (21 likes | 3 RTs) Read more β
ποΈ BUILD
DOOMQL: Yes, Someone Ran DOOM Through a Query Language. The latest entry in the noble "run DOOM in X" tradition, this time through query-language mechanics. Simon Willison covered it, which usually means there's a genuine technical insight underneath the fun β and there is: it's a creative stress test of in-database computation that's more instructive than most whitepapers. Read more β
vox-director: An Agent That Turns a Topic into a Finished Vox-Style Explainer Video. vox-director is an end-to-end agent that takes a single topic and produces a complete Vox-style paper-collage explainer video using Atlas Cloud + ffmpeg. It's one of the more complete automated video production pipelines to appear as open source β from research to script to visual assembly to final render. (103 likes | 10 RTs) Read more β
π MODEL LITERACY
Speculative Decoding (Multi-Token Prediction): Standard autoregressive generation produces one token at a time β each token requires a full forward pass through the model, which is why big models feel slow. Speculative decoding flips this: a small, fast "draft" model proposes several tokens at once, then the large "verifier" model checks all of them in a single forward pass, accepting the correct ones and rerolling from the first wrong token. You trade a little extra memory (for the draft model) for dramatically faster inference, because verification is parallel while generation is serial. This is exactly how llama.cpp's new Hy3 support runs massive MoE architectures locally β the multi-token prediction variant lets the draft model propose even more aggressively, which is why architectures that were cloud-only months ago are now running on local hardware.
β‘ QUICK LINKS
- headroom: Track Claude & Codex usage across multiple accounts without spending tokens. (73 likes | 7 RTs) Link
- Lorde vs. Ray-Ban Meta: Artist publicly calls AI glasses "not sexy" at their own sponsored festival β the cultural backlash product teams need to hear. Link
- BillAI Bass: AWS Strands Agents wired to a Big Mouth Billy Bass β a clean IoT + LLM integration reference disguised as a joke. (52 likes | 25 RTs) Link
- Claude as Mr. Meeseeks: Spawn an agent, do the job, vanish β ephemeral agent design as Rick and Morty intended. (87 likes | 28 RTs) Link
π― PICK OF THE DAY
Samsung's delete-or-consent ultimatum isn't just bad UX β it's a regulatory tripwire. When a company tells users "let us train AI on your health data or we'll delete everything," they're not offering a choice β they're leveraging data hostage dynamics against people who've been tracking their heart rate, sleep, and workouts for years. The real question isn't whether Samsung will back down (they probably will, partially). It's whether health-data regulators will treat AI training consent the same way they treat medical consent β because the answer sets the template for every fitness app, every wellness platform, and every wearable that's quietly collecting biometric data while waiting for the right moment to ask. GDPR's "freely given" consent standard was designed for exactly this scenario, and Samsung just handed regulators a textbook example of what coercive consent looks like. If you're building anything that touches health data, don't wait for the enforcement action β redesign your consent flows now, with a genuine "keep my data, skip the AI" option. The companies that get this right early will own trust when the regulatory hammer drops. Read more β
Until next time βοΈ