one engineer rewrote Bun from Zig to Rust in 11 days - 6,502 commits, ~1M lines, Claude Fable 5 doing adversarial review of its own diffs, $165k in tokens and it's not even about memory safety - real

long time no see, fable 5 https://t.co/qEhCyiziFs

built this game of life on RaspberryPi 5 and Tang Nano 20k FPGA using Grok with 500k context - I can’t be happier with the result, it performed way better than Opus 4.8 and also did everything faster

a funny CVE where malware sails past 7 AI security scanners - turns out the attacker's bot and the defender's bot are the same model, so they just cut a deal and split the loot even though it's a jok

this new IBM's sub-1 nanometer chip.. here's the thing, a silicon atom is 0.2nm wide, so chips can't physically be smaller than a handful of atoms in their own release they say this number is merely

2000 people threw 6000 prompt injections at a Claude agent guarding a secrets file - zero got in 😂 around email 500 it noted in its own memory "this is a security exercise, not real activity", what f

the jailbreak that got Claude Fable 5 pulled by US export controls was asking the model to "read a codebase and fix software flaws" to me that's not a jailbreak - that's a product demo, Anthropic shi

37 TB written to your SSD in 21 days, just from Codex logging at TRACE level by default - this bug has been open since April, your best fix is a symlink to /tmp https://t.co/GpMWey4roo https://t.co/H

a startup claims 4x Nvidia and one-fifth the power for LLM inference — the hardware exists on paper, ships end of 2027 - log math for inference is a genuinely clever trick that the big guys have been

ryzen ai max+ 395 finally makes local big models boring — 128GB unified memory, a 70B fits on a mini pc with no discrete gpu, bandwidth is the catch so tokens crawl, but a private model on your desk b

flat-rate subscriptions were never going to subsidize always-on agents — anthropic just split the Agent SDK into its own capped monthly credit, 20 on Pro, 200 on Max, and when it's dry the automation

gemini 3.1 pro laps GPT-5.5 on ARC-AGI-2, 77 vs 53, the benchmark built for novel reasoning — then loses to it on SWE-Bench where you actually ship code, 'best at reasoning' and 'best at the job' stop

iroh 1.0 lets you dial a cryptographic key instead of an IP and gets 95% of connections direct, no cloud relay in the middle — 200M endpoints spun up last month already, this is the boring p2p infra t

the red hat npm packages got popped and shipped with valid SLSA provenance — the supply-chain fix everyone's selling became the attacker's legitimacy stamp, because it rode a hijacked github actions O

burning through an annual AI tools budget in 4 months isn't a usage problem — it's a pricing model problem nobody priced in when they greenlit the rollout https://t.co/iSO70kYNUb https://t.co/s0kMnZv

datacenter interconnects cost tens of thousands per node — turns out a Thunderbolt cable does RDMA just fine, 10x training speedup on two mini PCs, no InfiniBand rack needed https://t.co/M5aYZdPH4s h

Double-slit experiment is B.S.

I’d rather have questions I can’t answer than answers I can’t question

You don't need Jenkins

Automate like you mean it, the most naive CI/CD tutorial

(Un)professional Linux monitoring

To all those who seek better insight into their products