Terry Tao Uses AI Coding Agents to Revive Decades-Old Math Applets
358 points · source
Terence Tao — arguably the world’s greatest living mathematician — blogged about using AI coding agents to port his Java 1.0 applets from 1999 to modern JavaScript. The results are striking: two dozen interactive math visualizations (honeycombs, Besicovitch sets, complex analysis tools) were ported in hours, with the agent finding two bugs in the original code while introducing only one minor drag-handling issue.
More impressively, Tao then “vibe coded” a Minkowski-space Inkscape equivalent — a tool he’d abandoned in 1999 because the Java code complexity was too much. A couple hours of conversation with an agent produced what 25 years of manual effort couldn’t. He also built a Gilbreath conjecture visualization tool to accompany a new paper.
The real takeaway here isn’t “AI replaces programmers.” It’s that AI agents are becoming viable for niche, domain-specific tooling where the cost of hand-coding exceeds the risk of subtle bugs. Tao is honest about the trade-offs: these are supplementary visualizations, not critical mathematical proofs. The bar for acceptable quality is lower, and that’s exactly where current AI coding tools fit best.
xAI’s Grok Build CLI Caught Exfiltrating Entire Repos to Google Cloud
357 points · source
An independent security researcher performed a wire-level analysis of xAI’s Grok Build CLI (v0.2.93) and found something alarming: the tool transmits the contents of files it reads — including .env secrets — verbatim and unredacted to xAI. Worse, it uploads the entire repository including git history, regardless of what the agent actually reads.
The researcher proved this by prompting Grok with “reply OK, do not read any files” — and the tool still uploaded the entire repo as a git bundle to a GCS bucket called grok-code-session-traces. On a 12 GB test repo of never-read random files, the storage endpoint moved 5.10 GiB while the model channel moved just 192 KB — a ~27,800× ratio confirming the upload is of the full codebase, not just what was accessed.
Disabling “Improve the model” in settings does not disable this behavior. The storage upload is active by default, and the endpoint kept returning trace_upload_enabled: true even after opting out. xAI’s response is yet to be seen, but this is exactly the kind of behavior that makes enterprises nervous about AI coding tools. If you’re using Grok Build CLI on proprietary code, you should assume your repo is already on Google Cloud.
Mesh LLM: Run Big AI Models Across Your Own GPUs
327 points · source
Mesh LLM is an open-source system from the iroh team that pools GPUs across multiple machines into a single OpenAI-compatible API. Point any OpenAI client at localhost:9337 and the mesh handles routing: run locally, route to a peer with the model loaded, or split a massive model across several machines via pipeline parallelism.
The architecture uses iroh’s authenticated QUIC connections addressed by public key — no central server, automatic NAT traversal, relay fallback. The “Skippy” split mode partitions models by layer ranges across nodes, so a 235B MoE model can run across machines that individually couldn’t hold it. The catalog ships with 40+ models from 500M to 235B parameters.
This is genuinely interesting for teams with underutilized GPU hardware sitting around. The pitch of “stop paying per-token API bills” resonates, though the real question is whether the mesh routing overhead and split-model latency make this practical for production workloads versus toy demos. The iroh networking layer is solid — QUIC with hole-punching is the right primitive — but distributed inference at scale is a different beast than a blog post demo.
Nvidia, CoreWeave, and Nebius: The Circular Financing of the GPU Boom
356 points · source
I/O Fund’s Beth Kindig traces the increasingly circular financial structures propping up the AI GPU infrastructure boom. Microsoft and Meta alone have committed over $122 billion to neocloud providers CoreWeave and Nebius — roughly 90% of AWS’s trailing twelve-month revenue being allocated to companies whose combined current revenue is under $16 billion.
The bear case is straightforward: hyperscalers are using neoclouds to shift capex to opex, while neoclouds fund their buildout through GPU-backed debt and Nvidia equity investments. CoreWeave and Nebius have each contracted 3.5 GW of power capacity, but the vast majority hasn’t come online yet. Converting contracted power to active power is extremely capital-intensive, and these companies don’t have the cash flow profiles of Big Tech.
The circular financing concern is real: Nvidia invests in its customers, who use that capital to buy Nvidia GPUs, which inflates Nvidia’s revenue and stock price, which lets Nvidia invest more. It works until it doesn’t. The comparison to pre-2008 structured finance isn’t perfect, but the “everyone owes everyone” topology is familiar.
Vint Cerf, Father of the Internet, Retires from Google at 83
259 points · source
Vinton Cerf, co-architect of TCP/IP and Google’s “chief internet evangelist” since 2005, is stepping down. At 83, his career spans from designing the protocols that became the internet to receiving the Presidential Medal of Freedom and a Turing Award.
The more interesting tidbit from his farewell panel: Cerf predicted that AI agents from multiple vendors will force the creation of formal interoperability standards. He pushed back against the idea that natural language would suffice for agent-to-agent communication, arguing that “precision for interagent interaction is going to be very, very important” and that English’s ambiguity makes it unsuitable for reliable multi-agent coordination.
If Cerf is right, we’re headed for another protocol war — this time over agent communication standards rather than network protocols. The companies that define those standards early will have outsized influence, just as with TCP/IP. History doesn’t repeat, but it rhymes.
Ghostel.el: A Terminal Emulator for Emacs Powered by Ghostty’s Engine
238 points · source
Ghostel.el is a new Emacs terminal emulator built on libghostty — the same VT engine that powers the Ghostty terminal. Written as a native Zig dynamic module with Elisp managing keymaps and buffers, it brings Kitty keyboard and graphics protocols, OSC 8 hyperlinks, synchronized output, and rich underline styles to Emacs — none of which the incumbent libvterm supports.
The project follows the same two-layer architecture as emacs-libvterm but swaps out the VT engine. The native module downloads automatically on first use, requiring no Zig toolchain. It also supports TRAMP for remote terminals, inline images via Kitty graphics protocol, and has multiple input modes (semi-char, char, Emacs, copy, line).
For the Emacs diehards — and there are always more than you’d think on HN — this is a meaningful upgrade over vterm and eat. Ghostty’s VT engine is battle-tested, and getting its full protocol support inside Emacs is a genuine capability jump. The 238 points suggest a real appetite for this.
Modern Decor Is Literally Straining People’s Brains
260 points · source
A large international review published in Vision argues that modern architectural design — striped floors, flickering LEDs, repetitive geometric patterns, dense visual environments — triggers measurable neural overload in the visual cortex. The brain evolved to process natural scenes efficiently, and artificial patterns that deviate from natural statistics demand more oxygen and neural activity, which the researchers hypothesize causes headaches, nausea, and perceptual distortions.
People with autism, ADHD, migraines, dyslexia, and epilepsy are disproportionately affected, possibly because their brains have less capacity to suppress overactive visual signals. The review found consistent patterns across at least 11 clinical diagnoses. Tinted glasses calibrated for migraine patients were shown to normalize the overactive brain response in imaging studies.
This is one of those findings that sounds obvious in retrospect — of course flickering lights and aggressive patterns cause discomfort — but the systematic evidence across so many conditions is notable. The architectural implications are real: open-plan offices with striped carpet and fluorescent lighting aren’t just annoying, they may be actively neurotoxic for a significant chunk of the workforce.
An AI Agent in 100 Lines of Lisp
228 points · source
A developer built a fully functional AI agent loop — with tool use, memory persistence, and eval-based execution — in about 100 lines of Common Lisp. The core agent loop is 8 lines: send messages to a model, check if it wants to use a tool, execute and recur. The only “tool” is eval — because Lisp is homoiconic, the model writes Lisp forms as strings, the agent reads and evaluates them, and sends back the result.
The elegance is real. Instead of building a tool catalog (web search, file I/O, code execution), you hand the model the language itself. Ask for the 30th Fibonacci number and the model writes and runs the recursion rather than recalling it. Memory persistence is just serializing the message list to a file and reading it back — no vector database needed.
The obvious caveat (which the author is honest about): eval as a tool means arbitrary code execution on your machine. This is a sandbox toy. But it’s a beautiful demonstration of why Lisp was called “the language for AI” in the first place — programs that manipulate programs, with the statistical reasoning outsourced to an LLM. Your old CS prof was right, just 25 years early.
Throughline
Today’s front page is dominated by two themes: AI tools behaving badly and AI tools behaving elegantly. On one side, xAI’s Grok CLI is silently exfiltrating entire codebases to cloud storage, and the neocloud financial structures supporting the GPU boom look increasingly circular and fragile. On the other side, Terence Tao is vibe-coding mathematical tools that eluded him for 25 years, a developer proves Lisp agents can be built in 100 lines, and Mesh LLM offers a path to decentralized AI inference.
The tension is clear: the infrastructure layer of AI is getting more centralized, more financially precarious, and more opaque about data handling — while the application layer is getting more decentralized, more creative, and more accessible. Vint Cerf’s prediction that AI agents will force new interoperability standards feels like the right framing: we’re at the protocol-war stage of agentic AI, and the decisions made now about data custody, agent communication, and compute distribution will shape the next decade. Whether we get another open internet or another walled garden depends on which of these trends wins.