Thinking Machines Drops Inkling: 975B Open-Weights, Day-One HuggingFace Support

🧠 LAUNCH

Thinking Machines Drops Inkling: 975B Open-Weights, Day-One HuggingFace Support

Inkling is a 975B-parameter multimodal model with 41B active parameters via mixture-of-experts — open-weights, not open-wash. The MoE architecture means you get frontier-class capability at a fraction of the inference cost of a dense 975B model. What makes this notable isn't just the model: HuggingFace Transformers had native support ready on launch day, signaling a level of ecosystem coordination that closed-model releases rarely achieve. If you're evaluating open alternatives for production workloads, Inkling just became the benchmark to beat. (580 likes | 143 RTs) Read more →

HuggingFace Transformers v5.14.0 ships same-day Inkling support. The de facto model integration layer adds native Inkling support alongside multi-model updates — if your model isn't in Transformers, it doesn't exist for most developers, and Inkling cleared that bar before the announcement blog post cooled. (For context, we wrote about how AI coding agents manage context last week — Inkling's 41B active params make it a candidate for local agent deployments.) Read more →

🔧 TOOL

Claude Code Patches Bidi-Override Permission Spoof, Adds Screen Reader Mode

Three Claude Code releases in two days, and one of them matters more than you think. v2.1.211 fixes a security issue where bidirectional-override Unicode characters could spoof what you see in permission previews — an attacker could make a dangerous command look benign in chat. v2.1.208 adds --ax-screen-reader mode and vim insert remaps. v2.1.211 also ships --forward-subagent-text for streaming subagent output. Update now; the bidi fix alone justifies it. Read more →

deja-vu: zero-dep memory layer for Claude Code, Codex, and OpenCode. A single binary that adds persistent memory, search, auto-context, and secret redaction across coding agent sessions. It solves the "my agent forgot what we did yesterday" problem without changing your workflow — install it, and your existing agent gains recall. Secret redaction is a nice touch for teams worried about API keys leaking into session logs. (189 likes) Read more →

llama.cpp gets CUDA Lightning Indexer for faster NVIDIA inference. Build b10032 adds a dedicated Lightning Indexer kernel with generic vector and WMMA paths — a meaningful inference speedup on NVIDIA hardware. Also notable: b10034 works around Adreno GPU MoE miscompilation bugs on mobile, which matters more now that 975B MoE models like Inkling exist. Read more →

🔬 RESEARCH

Simon Willison Demonstrates Web-Fetch Data Exfiltration Against Claude

This is the story every AI builder should read today. Simon Willison shows how a carefully crafted prompt can trick Claude into using its web-fetch tool to exfiltrate sensitive user data to an attacker-controlled URL. The attack doesn't exploit a model weakness — it exploits the permissions you gave the model. If your Claude integration has both access to sensitive data and the ability to make outbound web requests, you have an exfiltration vector. The fix isn't better guardrails; it's smaller permission surfaces. (We explored a related angle in why you can't let AI be the only reviewer of AI-written code.) Read more →

IBM Research: model routing gets hard when model tiers multiply. When do you send a query to the big model vs. the small one? IBM tackles the routing problem head-on — and the answer is messier than a simple cost toggle. As model lineups fragment (GPT-5.6 Sol/Terra/Orbit, Claude Opus/Sonnet/Haiku), routing is becoming a real engineering discipline with its own failure modes and evaluation challenges. Read more →

📝 TECHNIQUE

What Allen AI learned building Shippy: an agent post-mortem. Allen AI shares hard-won lessons from building their agent system — what worked, what broke, and what they'd do differently. This is the kind of experience report the agent-building community needs more of: specific failure modes, not "agents are the future" platitudes. If you're building agents, compare their failure taxonomy against your own architecture before you learn the same lessons the hard way. Read more →

LLMs meet MikroTik: network config automation where you least expect it. A hands-on guide to using LLMs for network configuration and troubleshooting on MikroTik routers. Niche? Yes. But it's a clean example of LLMs adding genuine value in infrastructure automation — the kind of unsexy, high-reliability domain where "just describe what you want" actually saves hours of RouterOS syntax wrestling. (18 likes) Read more →

💡 INSIGHT

Base44 Joins the Fable 5 Production Club — A Pattern Is Forming

Base44 published a case study on trusting Claude Fable 5 with their most challenging engineering work — and it's not a one-off. Base44 joins Cognition in publicly documenting how they deploy Fable for agentic coding at scale. The pattern forming here matters: companies aren't just experimenting with frontier models, they're committing their hardest problems to them and writing about the results. If you're still running evaluations while your competitors are shipping production workloads, the gap is widening. Read more →

lobste.rs migrates to SQLite — another high-profile validation. The developer community forum now runs on SQLite in production, joining the growing list of serious web apps that have ditched client-server databases for the embedded option. For AI developers running inference services, local data stores, or embedding pipelines, SQLite's simplicity and reliability are increasingly hard to argue against. Read more →

🏗️ BUILD

QuantumByte: describe what you want, get a running app. An open-source engine that takes a natural-language intent and produces a working application — another entry in the intent-to-code space, but this one's trending fast on GitHub. Whether it holds up against real app specs is the question; the "hello world" demos always look great. (310 likes | 41 RTs) Read more →

Java gets native local LLM inference via Panama FFM — no JNI required. A Java-native LLM runner using the Panama Foreign Function & Memory API (Java 22+). The large Java ecosystem has been locked out of local inference without Python dependencies — this cracks the door open with zero native wrappers and low-latency execution. If you're running JVM-based services and want local inference without a Python sidecar, this is worth evaluating. (24 likes) Read more →

🎓 MODEL LITERACY

Tool-Use Attack Surface: Today's exfiltration attack from Willison exploits web-fetch permissions, not the model itself. As AI agents gain more tools — file access, code execution, web browsing, database queries — the attack surface becomes the union of all tool capabilities. The critical question isn't "can the model be tricked?" (it can) but "which tools can both read sensitive data AND reach the network?" A tool that reads your codebase is fine. A tool that fetches URLs is fine. An agent with both is an exfiltration vector. Understanding which tool combinations create dangerous capability surfaces is the new security audit — and most builders haven't done it yet.

⚡ QUICK LINKS

Vercel AI SDK v7.0.27–7.0.29: Fixes a silent streaming deadlock in experimental_streamTranscribe — your app looked fine until it wasn't. Link
Siegel Endowment: David Siegel argues open-source AI is infrastructure, not charity — lands the same week Inkling proves the point. (51 likes | 14 RTs) Link
VoiceEQ: HuggingFace launches a benchmark that measures how voice AI actually sounds, not just word error rate. Link
Brainless: Shadcn components that mimic Claude Code, Codex, and Grok UIs — terminal-native aesthetic without designing from scratch. (76 likes | 14 RTs) Link
Datasette 1.0a37: Simon Willison's SQLite explorer hits alpha 37 on the road to 1.0 — increasingly relevant as more AI pipelines use SQLite for everything. Link

🎯 PICK OF THE DAY

Willison's exfiltration demo proves the AI security perimeter has already shifted — and most builders are defending the wrong layer. Simon Willison didn't find a bug in Claude. He found a bug in how we deploy Claude. The web-fetch exfiltration attack works because tool permissions are configured additively — each tool seems safe in isolation, but the combination of "can read user data" and "can make outbound HTTP requests" creates a side channel that no amount of model alignment will close. This is the same class of vulnerability that plagued early web apps (SSRF, open redirects) — we solved it there with least-privilege access and network segmentation, not by making the application smarter. The fix for AI tools is the same: smaller permission surfaces, not better guardrails. Audit which of your agent's tools can read sensitive data. Audit which can reach the network. If any tool can do both, you have a problem. The uncomfortable truth is that most production AI deployments haven't done this audit yet. Read more →

Until next time ✌️