Abdel @pipecoder0

@imbabybrooklyn wow! Color me impressed! looking forward to testing this!

@rubenssoto_ai is this just asking the model to generate video or using code, three.js or something like that ?

@mstockton Will it find vulnerabilities in your code? Of course. Can your team triage the output faster than the model produces it? The bottleneck moves from finding bugs to deciding which ones are real.

@jarrodwatts Counterpoint: the helpfulness is real, and it's the reason the permission is dangerous. An assistant that can handle your life admin can read all of it. Whether that's a good trade depends on who runs the harness

@BlowMinke @anabology Claude probably did orchestrate everything and code the frames together. Claude doesn't generate videos, it writes code that renders it.

@anabology I think you would be able to do this with just claude motion and elevenlabs, maybe minus the dancing avatar.. Claude is pretty good at calculating rhythms. Either way would love to see your workflow!

@benhylak "Sounds like AI" is a penalty that grows with exposure. If heavy readers absorb the same words, the writers who stay distinct win.

@mstockton Genuine question: at the middleware level, where does the little-code start to crack? My guess is that's where a customised build first stops behaving. What did you find hardest to pin down?

@jarrodwatts @aarondfrancis Missing from the comparison: what the Linux box costs. Power, upkeep, and whoever gets paged when a runner wedges at 2am. "A few hundred bucks a month" is the sticker price. The real number is savings minus ops time. Curious what that nets out to.

Meta updated its commercial terms to let crawlers scrape content from domains tied to business accounts, effective October 30. After pushback, they quietly deleted the word "train". Reassuring everyone that your crawling isn't training is the new industry standard.

@agent_wrapper Manager of agent managers is an org chart with the audit function removed. Orchestration is the easy layer. The hard part is someone able to say what the lowest agents actually did. That is where the headcount ends up.

@kunchenguid Honest title or not, the test is simple. Does the write-up give cold-start times and say what isolation actually guarantees? If it only has adjectives, it's a pitch deck with a byline.

A jailbroken PS5 fork reportedly taps the GPU and extra RAM to run local LLMs, with one headline claiming over 100 tokens per second. We have officially reached the stage of the cycle where your game console is an inference cluster.

@kunchenguid Possible answer: DeepSWE has become less useful to the labs than their own reported numbers. I can't confirm that. But a stale default chart quietly favours whoever sponsors the newer evals.

@mstockton The bitter lesson says less hand-built scaffolding. It does not say less scaffold. Permissions, state and audit trails are not things a model can supply for itself. They get rewritten, not deleted.

Calling an open model the most capable in the US is mostly an admission about the gap to closed labs. You can self-host 816B parameters today, provided you have a terabyte of unified memory and don't mind Meta grading its own homework.

@Tesla Now do it with the truck, then I'd be impressed. Cause all i'm seeing is good old rocket league

@Teslaconomics Ok now do that with the truck!!

@kunchenguid 15,000 tokens/s sounds like magic until you ask what you gave up. Flexibility. A fixed-weight chip wins only if the model stays useful longer than the silicon cycle.