AI auto-formalizes counterexamples before humans

Jane Street released Incremental, an OCaml library that lets you build self‑adjusting computations that update efficiently as inputs change. It shines for spreadsheet‑style calculations, live GUI views, and keeping derived data in sync, saving time and memory in data‑heavy apps.

Jelly UI brings soft-body physics to native HTML form controls via a dependency‑free Web Components library. The jiggly interactions add tactile feedback without sacrificing accessibility, dark‑mode or RTL support, letting developers create playful, yet standards‑compliant interfaces.

Large language models are now generating and auto‑formalizing mathematical counterexamples faster than human experts. Systems like Logical Intelligence's autoformalizer and OpenAI's Sol have turned raw AI‑written proofs into Lean code within days, reshaping how mathematicians verify results.

In a live fireside chat, Anthropic’s Cat Wu and Thariq Shihipar unpack Claude Code’s latest updates, Claude Tag, the Fable model, and hardened agent security. They also share real‑world evaluation practices and tool design lessons that will shape how developers integrate AI coding assistants into production pipelines.

At an AI Engineer World’s Fair fireside chat, Anthropic’s Cat Wu and Thariq Shihipar detail Claude Code’s latest upgrades, Claude Tag now powers 65% of internal PRs, Fable’s prompt size cuts 80%, and automated code review handles outer layers. They also explain security review shifts and why “don’t do X” lists hurt modern models, reshaping how developers adopt AI coding agents.

Stencil shows that a “plan‑then‑execute” chain with a frontier model and a cheap executor actually costs more, because the expensive model reads the whole codebase twice. By limiting the frontier LLM to a single, decisive edit, you cut token usage and spend ~14% less while keeping the same success rate.

A new design lets an AI agent sit alongside you inside an app, editing the same document you’re working on in real time. In a music‑making demo, the agent changes chords, adds bass lines, or transposes a song while you keep tweaking. This flips the usual chat‑box model, turning AI into a collaborative co‑author for any content.

Cursor’s latest swarm, guided by planner agents and cheap worker models, built SQLite from its docs and passed 80% of a held‑out SQL test suite in four hours, far outpacing its predecessor. The new architecture lets the spec drive the swarm, keeping planning context light and execution cheap, promising scalable AI‑assisted development.

In an email to OpenAI’s board, Sam Altman revealed the company is planning an open‑source language model with roughly GPT‑3‑level capability that can run on consumer hardware. He said the move aims to undercut rivals and make it harder for new, high‑cost AI ventures to attract funding. The proposal will be discussed at the next board meeting.

Simon Eskildsen shows that MySQL’s transaction ceiling isn’t the OS’s fsync rate; the database batches writes and can push far beyond the 1 kHz fsync limit. The deep dive quantifies how hardware, indexing and write size shift the realistic 5k‑15k TPS sweet spot.

The napkin-math GitHub repo bundles memorized latency and throughput numbers for common operations, from memory reads to network speeds. Engineers can plug these figures into quick back‑of‑the‑envelope calculations, slashing the time spent hunting benchmarks and improving capacity planning accuracy.

France's ANSSI will stop certifying security products that lack quantum‑resistant encryption from 2027. Losing certification means exclusion from the French government market, pushing vendors to adopt PQC now or lose a major European foothold.

LLM tools have collapsed the cost of writing code, invalidating many traditional management rules that tie headcount to capacity. Managers should audit each practice for assumptions that rely on code cost, keeping only those that rest on human coordination, trust and verification. The shift lets teams stay lean while preserving essential friction.

Z.ai has finished a 1‑gigawatt AI data center built entirely with domestically produced chips, and parts are already running. The move gives China a large‑scale compute platform independent of U.S.-controlled Nvidia silicon, accelerating its push for AI self-sufficiency.

Martin Fowler distills the second Future of Software Development Retreat into five headline findings: verification now outpaces code generation, a new "harness engineering" discipline emerges, apprenticeship pipelines are collapsing, executive‑engineer expectation gaps pose major risk, and legacy modernization offers the clearest near‑term ROI. The notes also flag LLM misuse and citizen‑developer security concerns.