Dwayne Samuels

The easiest way for agents to use the browser👀

self-healing selectors are the "it's fine" dog meme of test infrastructure the DOM is on fire. your selector melted. the healer pointed at a different element. your test passed everything is fine

the pitch for self-healing selectors: "your tests won't break when the UI changes" what they actually optimize for: test pass rate what they ignore: whether the test is still checking the right thing that mismatch is the entire problem think of it as a signaling system: without healing. DOM changes, test fails. honest signal. expensive, but informative with healing. DOM changes, test passes. ambiguous signal. cheap, but dangerous you are trading false negatives for false positives. and false positives are how QA systems rot from the inside "closest match" is not "correct match." the old Add to Cart button and the new Subscribe button can look identical to a heuristic engine. same position, same classes, similar role. the healer picks it. your test passes. your metrics drift until someone notices the deeper issue: healing has no concept of intent it matches structure, not meaning your test is asserting: "this is the Add to Cart action in this purchase flow" the healer is asserting: "this looks similar enough" those are not the same system now. self-healing is not universally bad. it works on low-risk smoke tests. it works when changes are logged and reviewed. it works when strong selectors already exist underneath it fails when it becomes a substitute for selector quality. when it optimizes for developer comfort instead of system truth what actually works: know which selectors are fragile before they break. score them. penalize positional paths and dynamic classes. treat elements as objects with identity, not just DOM paths and when something breaks, treat that as information. not noise to suppress

A breakthrough in real-time video generation. As a research preview developed with @NVIDIA and shared at @NVIDIAGTC this week, we trained a new real-time video model running on Vera Rubin. HD videos generate instantly, with time-to-first-frame under 100ms. Unlocking an entirely new creative paradigm and bolstering the foundations of our General World Model, GWM-1. Real-time generation opens a fundamentally different design space for video models and world simulation. We're investing in co-designing our models alongside advances in hardware to keep pushing this frontier.

self-healing selectors don't fix broken tests. they remove the signal that tells you your test is broken your selector broke. the tool scanned the DOM, found something "close enough," and swapped in a new path. no review. no approval. it just… healed it's basically autocorrect for your test infrastructure... and we all know how "accurate" autocorrect can be

5.4-mini is roughly Sonnet 4.6 intelligence but 70% cheaper and like 3x faster

when bro's pr has -- and emojis

They're building OpenClaw

we built the Jira integration because we kept hearing the same thing: "the ticket was perfect when i wrote it, but the engineer still had to call me" the ticket was not perfect, it was complete in PM language and empty in engineer language. those are different completeness standards and most teams do not realize they are using the wrong one

Introducing Humanoid Atlas, the Bloomberg Terminal for humanoids. Every OEM, every supplier, every dependency humanoids.fyi

a CRO strategist told me she spends more time explaining her findings to engineers than she spent finding them the analysis took an hour. the Slack thread explaining which element she meant took two hours and a screen share. the problem was never her analysis, it was the format it traveled in

@kimmonismus either exponential efficiency or exponential energy requirements, no in between haha

1) what?

every QA team i talk to has the same complaint: they do rigorous work and have nothing to show for it at the end of the sprint the test passed. the verification happened. but the evidence lives in someone's memory and a browser tab that got closed. if nobody was watching, it might as well not have happened

We just crossed $5M ARR fully bootstrapped Introducing CrowdReply 2.0 The new benchmark of ranking in AI Answers

the best product decisions we made were not adding features they were removing steps between "i see the problem" and the engineer can start fixing it. stop sending Loom videos engineers cannot act on. send a technical artifact, the exact CSS, DOM state, and console errors from the moment it happened

The power density of Vera Rubin is incredible, especially in this form factor. This has a triple benefit for space: - Ruggedization against launch vibration is substantially easier for smaller boards - Volume is hugely constrained on orbital datacenters - Radiation shielding is less massive for smaller payloads If ever you needed external validation that space datacenters are happening, this is it.

the pitch wins the contract. the handoff decides whether it renews i keep noticing how much investment goes into one and how little goes into the other

the role transition chain on a typical CRO experiment: the strategist identifies the opportunity. they know the page, they know the hypothesis, they have the data. they record a Loom, point at the region, explain the thinking. they are done with their part and it was genuinely good work the designer receives the Loom. builds the variant in Figma. clean comp, approved by the client. their part is done too the developer receives the Figma comp and the Loom. now they need to answer a question nobody upstream was responsible for: which exact DOM element receives this treatment? what is the selector? what state does the page need to be in? is the element inside an iframe? does it render the same way across viewports? none of that was in the Loom. none of it was in the Figma file. the strategist had no reason to capture it. the designer had no reason to know it. and the developer has no way to answer it without going back upstream or guessing when the developer guesses right, nobody notices the gap existed. when they guess wrong, the experiment runs for three weeks on the wrong target and the postmortem blames communication it was not communication. it was a missing artifact at a role transition. the handoff crossed a domain boundary and nothing carried the context across

They’re building OpenClaw

a strategist should not have to know what a CSS selector is but the process should not let the experiment move forward without one either the gap between those two facts is where i keep watching teams lose weeks