Excited to share more about what I have been working on!.

RT @zhenthebuilder: Build with any framework and languages on Replit Agent is fun, this is what Agent used to write code for itself.

RT @fcvprzhfgsybj: Hiring for new role at @ExaLaboratories!. If you have a strong background in chip design (RTL, & + if mixed) with tapeo….

getting hit with the “welcome back” at the border is one of the best feelings ever.

i think its under-appreciated how modal is the perfect swiss knife for doing adhoc research-adjacent work when you're operating without well-built lab infra. have used it heavily for rapid eval creation and benchmarking runs with oss models, synthetic data generation, one-off.

RT @giansegato: published a new essay on my blog! "Building AI products in the Probabilistic Era". admittedly it's a very long one (took mo….

finding it to be a really subpar UX for a large distribution of use cases. lowering reasoning effort is a no go because you just introduce another knob in my head to think about to get what i want, and guess what people often love dialing knobs up when they are uncertain.

one of the og lessons building coding copilots and then coding agents is that response latency really matters, especially used in an inner loop work context. gpt5 (and variants) are way too slow to the point i'd rather using something else for "daily driving", e.g. Claude. of.

this brother's been heads down cooking.

Time to First Useful Token and e2e Time to Solution are undeniably important metrics. The model can do all the thinking the model wants to get the best possible OR correct answer but most if those numbers are not constrained for “inner loop” products, you’re going to bleed users.

literally every search i start from the chrome url bar becomes useless. bleeding users. if they want to compete with openai, this and 20 other clearly obvious things that I have found need to be fixed.

google's AI Overview product team should really go talk to their users. if you do a search and get an AI overview response, if you click the (extremely poorly placed) "Dive deep in AI Mode" button, then it completely restarts the conversation, possibly wiping useful info. ffs.

seeing a trend in startups increasingly hiring young ops/growth talent coming out of blackstone, blackrock, etc. consulting was always common and is only continuing to grow, but interesting to see expansion in this direction.

telling these waterloo youngins about ulmfit, sentiment neuron, and the unreasonable effectiveness of rnns. they're cracked but they gotta know their history.

business / use-case specific rl finetuning, synthetic data, airgapped on-prem deployments are the obvious (and good) plays. thinking more about what new things get fundamentally unlocked. we had qwen etc. already, but on-device readiness is a meaningful differentiator.

RT @DimitrisPapail: Excited about our new work: .Language models develop computational circuits that are reusable AND TRANSFER across task….

RT @sidbidasaria: Claude Code is getting a brand new feature: custom subagents. Type `/agents` to get started.

RT @thekaransinghal: 📣 Excited to share our real-world study of an LLM clinical copilot, a collab between @OpenAI and @PendaHealth. Across….

every "agentic search" pipeline is basically a dynamic retrieval pipeline built on the fly as a function of the query. similarly, every "/compact" style context compaction pipeline should be a function of the next task. pattern probably applies in more use cases.