Microsoft Drops Claude Code; the Case for Using AI to Code Slower

Huawei board member He Tingbo unveiled a new chip design framework at IEEE ISCAS 2026 in Shanghai, claiming the company can reach 1.4nm-class transistors by 2031 with 55% higher density, all without the EUV lithography machines US sanctions block them from buying. The firm says it spent six years secretly refining the methodology, designing 381 chips based on the principle. Huawei also introduced a 'Tau Scaling Law' it positions as a successor to Moore's Law.

Pope Leo XIV used his first encyclical, Magnifica Humanitas, to call for governments to slow down AI development and build robust legal frameworks with independent oversight. Anthropic co-founder Chris Olah attended the Vatican launch. The document urges that AI data ownership not remain solely in private hands, calls for protection of workers and children, and warns that autonomous weapons have advanced 'practically beyond any human reach to govern them.' The encyclical also repudiates the Church's traditional just war doctrine.

Kevin O'Leary's Stratos Project would cover 40,000 acres of Utah's Hansel Valley, more than twice the size of Manhattan, and consume 9GW of power, nearly double the state's 2025 peak electricity demand. Box Elder County commissioners approved the project, which O'Leary positions as a way to establish American AI dominance. Residents and experts warn of environmental damage and strain on overtaxed water supplies. The first phase alone is projected to cost more than $4 billion.

Microsoft is migrating engineers off Claude Code to GitHub Copilot CLI by June 30, citing toolchain unification. The real driver is economics: Uber burned its entire 2026 AI coding budget in four months after usage jumped from 32% to 84% of engineering, with individual engineers spending $500 to $2,000 monthly on tokens. Anthropic itself banned the OpenClaw framework from consumer plans after single instances consumed up to $5,000 daily. The era of unlimited per-seat AI coding access is ending, replaced by metered utility billing.

ClickHouse's engineering team breaks their AI adoption into three levels: chat copy-paste (obsolete), agents in IDE and CLI (daily work), and autonomous multi-agent loops (still maturing). Early Claude Code got lost in their C++ codebase. Opus 4.5 changed that. Agents now resolve merge conflicts better than humans in nearly 100% of cases, and the 'agent writes, you review' pattern produces higher quality than typing code yourself. The biggest lesson: agents that read well write well.

CockroachDB's engineering team wrote down their architectural constraints for vector search: no central coordinator, no large in-memory caches, real-time updates, no hot spots, sharding support. Most popular algorithms failed at least one requirement. Their response was C-SPANN, an index that treats vector data as ordinary table rows inside CockroachDB rather than a separate system. ByteByteGo walks through the design decisions, including how they handle the curse of dimensionality in a distributed setting.

The Redis creator explores what happens when you connect multiple Mac machines for distributed inference. A single Mac Studio M3 Ultra runs DeepSeek V4 PRO at 150 tokens/second prefill. DwarfStar splits transformer layers across machines and uses micro-batching to increase throughput without duplicating memory. With Apple's M5 Max at $6-7k per machine, the economics start competing with GPU rentals for certain workloads. The project is early but the hardware trajectory favours it.

Lawson runs multiple independent agents (Claude, Codex, Cursor Bugbot) in parallel sweeps to hunt bugs, categorises findings by severity, then filters false positives manually. The approach regularly surfaces pre-existing issues the team missed. His thesis: AI is most valuable when it amplifies careful, quality-obsessed engineering rather than replacing it with velocity. Clearing context between model sweeps minimises hallucinations and keeps each pass independent.

Neevash Vakharia traces a pattern: companies whose products are increasingly consumed by software rather than people are pulling ahead. Turbopuffer crossed $100M ARR in 19 months. Modal closed a $355M Series C. Mintlify's data shows agents now account for nearly half of all traffic across documentation sites. Meanwhile, companies that sell interfaces to humans face a squeeze as capable agent loops, run cheaply in parallel, can recreate most of what the application layer currently offers.

The Turing Award winner and former Meta Chief AI Scientist unpacks his thesis that LLMs cannot plan, predict consequences, or model the physical world. He explains JEPA, his alternative architecture that learns abstract representations rather than generating pixel-level predictions. LeCun reveals the real story behind his departure from Meta (he had zero technical influence on Llama), the genesis of his Tapestry project for sovereign open-source AI, and predicts OpenAI is 'the next Sun Microsystems.'

Alexander Volchek tested what happens when you bypass the CRM entirely and hand raw deal data to an LLM. The model returned a 'large, ugly, useful canvas' of patterns, strange cuts, and follow-up questions that a senior analyst's summary did not surface. His argument: AI does not need to replace the person. It can replace the interface that sits between the person and the data. Every SaaS product whose value is organising information rather than generating it is exposed.

Mario Zechner and Armin Ronacher, two engineers behind core pieces of OpenClaw, warned the Wall Street Journal that AI tools are flooding codebases with 'vibe slop.' InfoWorld's Matt Asay makes the case that AI coding is powerful the same way power tools are powerful: they help skilled people do more, and help unskilled people make bigger mistakes with greater confidence. The trouble is not obviously broken code but plausible output that slips into production.

Alejandro Piad Morffis examines four cases of AI-assisted discovery and finds a consistent four-role loop: poser, proposer, verifier, curator. AI excels in the proposer slot. Lean's type-checker, AlphaFold's crystallography, and Berkeley's A-Lab robotics serve as verifiers. His diagnostic for evaluating any AI discovery claim: what was the verifier, and who built it? The institutional bottleneck is not better proposers but stronger verification infrastructure.

Johanna Larsson argues that using LLMs to write breaks an implicit contract: the writer should invest more intellectual effort than the reader. She cites Oxide Computers' RFD 576 and a Max Planck study showing ChatGPT vocabulary patterns (delve, realm, meticulous) seeping into spoken language. When students get better grades submitting AI work than genuine effort, the incentive structure is broken. The piece is a commitment to writing without LLM assistance.

Lance is a unified multimodal model from ByteDance Research that supports image and video understanding, generation, and editing within a single architecture. At 3B active parameters, it delivers competitive performance across image generation, editing, and video benchmarks. The model was trained entirely from scratch on a 128-A100 GPU budget using a staged multi-task recipe, making it one of the most efficient unified models at this scale. Apache 2.0 licensed.

Claw Patrol parses agent traffic at the wire level and gates each action against rules written in HCL. Block destructive SQL, pause kubectl delete until a human approves, deny secret access from the cluster. Conditions are CEL expressions over protocol-level facts: SQL verbs and table names for Postgres, resource and verb for Kubernetes, method and path for HTTP. Three deployment modes: standalone gateway, host-wide WireGuard tunnel, or per-process isolation via network namespaces.

A single skill file that stops AI coding tools from generating boring, generic output. Install it in Claude Code, Cursor, or Codex and it rewrites the model's default aesthetic: no placeholder copy, no generic component names, no cookie-cutter layouts. At 19,579 stars and 542 new ones today, it is the fastest-trending GitHub repo in the AI tooling space this week. The approach is simple: a set of constraints that raise the floor on what AI-generated code looks like.

Starting in MLflow 3.12.0, two environment variables route every Claude Code session through the AI Gateway. Each request becomes a traceable, governable event with budget enforcement and guardrails, without touching application code. Developers keep their own Anthropic credentials while the gateway adds governance on top. Given that Microsoft just cited runaway token costs as the reason for dropping Claude Code, this is timely infrastructure for any team trying to keep agentic AI spend under control.

Webwright rethinks web automation by giving agents a terminal instead of a persistent browser session. The agent launches disposable browser instances, writes Playwright scripts, and persists artefacts in a local workspace. The harness is roughly 1,000 lines across three modules. On benchmarks it hits 86.7% accuracy on 300 live web tasks and scores 60.8% on the Odysseys long-horizon benchmark, a 35% relative improvement over prior state of the art. Failed tasks leave scripts and logs you can actually debug.