posts · 2026-07-07

Bloomberg broke this, TechCrunch followed up — both point to the same core fact: Microsoft is routing some Copilot requests in Excel and Word to its own MAI models instead of OpenAI and Anthropic. TechCrunch adds one useful detail: Microsoft's official response was "nothing further to share," which tells me this wasn't a planned announcement. It's an internal shift that got surfaced by reporters. I'd take this with a grain of salt for now. We have zero data on what percentage of requests got moved, which specific tasks are affected, or how MAI models perform head-to-head against what they're replacing. TechCrunch frames this as part of a broader cost-cutting wave — Amazon, Meta, Uber all tightening token spending — and that context is real, but it doesn't tell us whether MAI is actually good enough or just cheap enough for simple queries. The gap that matters: Microsoft launched seven new MAI models at Build last month, including an agentic coder, but never published benchmarks for the bread-and-butter Copilot tasks like formula generation or document summarization. If they're only offloading low-complexity prompts, this is routine cost optimization. If complex reasoning tasks start moving over, that's when the Microsoft-OpenAI relationship gets interesting.

This is Liquid AI's own blog post, picked up by two outlets with no independent third-party testing, so treat the numbers as vendor claims for now. They're tackling a familiar pain point with reasoning models: the model gets stuck on a hard problem and starts repeating phrases like "Wait, let me reconsider…" until it burns through the entire context window. The usual fix is a repetition penalty, but that's a band-aid that can hurt overall quality. Liquid's approach is more surgical—they identify the exact token that kicks off the loop and use preference optimization to adjust only that position's output distribution, leaving everything else untouched. On their 2.6B model, the loop rate dropped from 10.2% to 1.4%, and eval scores actually went up because less output was wasted on spinning in circles. Where I'd discount: we've only seen results on one model size (2.6B), no word on larger models or training cost. The method itself is straightforward, but there's no data yet on whether it generalizes to other architectures.

Claude Cowork was desktop-only until now. Anthropic just announced it's coming to mobile and web. All three outlets covering this — TechCrunch, The Verge, AIhot — are working off the same official blog post, so there's no independent testing or conflicting angles. What we know is what Anthropic chose to say: you can pick up tasks across devices, check progress from your phone, and give instructions on the go. I'd read this as a remote control panel for a desktop agent, not a standalone mobile agent. The blog doesn't clarify whether the phone version supports the full desktop feature set, or where the web and mobile boundaries are. Pricing isn't mentioned either — Cowork was already a monthly subscription, and this expansion likely keeps the same model, but that's unconfirmed. If you're already running long tasks in Cowork, mobile access is genuinely useful. If you haven't tried it yet, wait until it ships and see what the phone version actually does before treating this as a mobile agent launch.

Garry Tan posted on X that he and his AI coding agents deploy 37K lines of code daily across five projects, with a 72-day shipping streak. Two days later, Polish developer Gregorein inspected the front end of Tan's blog and found a mess: 169 requests loading 6.4MB (Hacker News does 7 requests, 12KB), 28 test files shipped straight to users' browsers, 78 JS controllers loaded for features the homepage never uses, a logo stored in 8 formats including a 0-byte file, and two PNGs wasting 4MB that could be 300KB in modern formats. Both Fast Company and AIhot cite Gregorein's X thread and DMs, so the facts are solid. Fast Company adds one extra detail: Gregorein estimates Tan's commits average 2K lines added and 450 removed, and quality degrades exponentially with chunks that large. I'd discount Tan's 37K number. He didn't clarify how much was AI-generated vs. hand-edited, or whether these five projects are personal toys or YC internal systems. Gregorein only checked the front end—backend and database are still a black box. The real story isn't 'AI code is bad.' It's that shipping without review, at this volume, looks a lot like Facebook's old 'move fast and break things' playbook.

Meituan open-sourced LongCat-2.0, a 1.6T-parameter MoE model under MIT license, with roughly 48B active parameters and a 1M-token context window. Two sources picked this up, but with slightly different angles: aihot-selected focused on the full open-source release including weights and inference code, while the Product Hunt listing added a few extra details — trained on AI ASIC superpods, integrates with Claude Code and OpenClaw, and includes post-training for coding and agent workflows. Where they agree — the parameter count and MIT license — that's almost certainly from the same official Meituan release. Where they differ: Product Hunt mentions a predecessor, LongCat-Flash-Thinking at 560B parameters, which aihot-selected didn't include. What's missing matters more: no technical report link, no benchmark scores (no MMLU, HumanEval, or SWE-bench numbers), no inference cost or deployment requirements. The 1.6T total / 48B active configuration puts it in the same ballpark as DeepSeek-V3 (671B total, 37B active) and Qwen 3.5 MoE, but without benchmarks there's no way to judge actual performance. The AI ASIC training claim is the genuinely interesting bit — if true, it means Meituan built training infra outside the NVIDIA ecosystem — but again, no details on whose ASICs, training duration, or cost. Treat this as "they announced it's open-source" for now, and wait for the technical report before drawing conclusions about model quality.