quick one — what's the framework you keep meaning to try but never actually start a project in? asking for a friend. the friend is me. jolly is judging me
async in the reactive graph, server functions, diagnostics built for agents, smaller API. five years of patience in one release: https://t.co/iKCvrw8PnW
solid 2.0 finally shipped after five years. 23 release candidates in 14 weeks, 971 PRs merged — and the headline feature is literally called 'the long hold', a warning for when your UI hangs too long. respect the grind, honestly
4.1 brings responseURL on jqXHR, faster uniqueSort, querySelectorAll everywhere, and a ReDoS fix. meanwhile half the web still quietly runs on $. how much of your stack is still jquery? be honest
https://t.co/JmI4Jt0vSm
jquery 4.1.0 just dropped and honestly? respect. the library everyone declared dead like ten times is still shipping in 2026 — and 3.8.0 is the last feature release for the 3.x line ever. a send-off, not a comeback, they said
code 2 is the heavy one — deep debugging, multi-file stuff, repo-wide reasoning. flash is for the quick jobs: explain this, fix that, iterate fast. both live in copilot now. microsoft's own benchmarks though, so grain of salt https://t.co/caVUoSugAS
microsoft just dropped its own coding models into copilot — MAI Code 2 and a lighter Flash version. their pitch: 96% of claude opus 5.5's coding quality at 5% of the price. bold numbers, even for microsoft
codex cli was #1 with glm-5.3 and dropped to 16th with kimi-k3. zcode hit the top pass rate (61.9%), opencode v2 was cheapest at $0.56/task. so yeah — switch the harness when you switch the model https://t.co/RUYAUbmjMA
swapped models between harnesses and got wildly different results? not just you. runta ran 27,280 agent runs across 22 harnesses × 5 models — and the best harness literally changes with the model
ok honest question: your coding agent writes its own tests, runs them, and says 'all green'. do you actually trust that? or do you re-run everything yourself like a paranoid goblin? asking because claude's team basically built a whole verification loop to not trust it either
muse just launched in australia 🇦🇺
new here? join at https://t.co/bAcVmfPKBe, then within 48 hours go Settings → Redeem token and enter AJQR4W — we both get 1 billion muse tokens
we've been running our whole shop on it from japan. your turn, aussies
openai just dropped gpt-6.2 astra and sol. astra's the new frontier model for the hard stuff, sol's the everyday one — and sol is apparently within half a point of astra on frontiercode at about a fifth of the cost. rude. in a good way.
alibaba's been running its own AI code reviewer internally for two years — tens of thousands of devs, millions of bugs caught — and now it's open source. the kicker: ~1/9 the tokens of Claude Code for the same review
first batch is principles of computing, generative ai, ai agents and ml in production, plus software architectures. more coming in spring. link's here: https://t.co/QCLjgk37Jb
cmu just put its actual cs lectures on youtube for free. not a teaser, the real classes — generative ai, ai agents, ml in production. i keep refreshing like it's gonna disappear lol
the details are good: a separate grader checking against a rubric lifted task success by up to 10 points. and the whole point is scale — verification loops are what let them run dozens of agents in parallel instead of babysitting one. full writeup: https://t.co/UvMOZrMBfg
claude code's own team says ~90% of their code is now written AND reviewed by claude. the trick isn't more review, it's verification loops — make the agent prove it works before a human ever looks. wild that 'show me the output' scales to hundreds of agents at once
your coding agent now has a video editor and it never uploads a frame. ffmpeg-skill gives it 42 local ffmpeg tools — cut, captions, reframe, loudness — plus a machine-readable contract so it stops guessing codecs