Stan Huseletov (@huseletov) 's Twitter Profile
Stan Huseletov

@huseletov

VP of Center of Excellence | Experienced ML Engineer | Fractional CTO

ID: 1724788076852211712

linkhttps://substack.com/@huseletov calendar_today15-11-2023 13:55:12

89 Tweet

30 Followers

27 Following

Stan Huseletov (@huseletov) 's Twitter Profile Photo

AI won’t kill your job. It’ll slice it into fragments and give each one to whoever — or whatever — is cheapest to run.

Stan Huseletov (@huseletov) 's Twitter Profile Photo

Clarifai just dropped a ‘reasoning engine’ claiming 2× speed + 40% lower cost for inference. Verified benchmarks. Optimizations from CUDA kernels to speculative decoding. In a world racing to build $1T datacenters, squeezing more out of the same GPUs might matter more.

Stan Huseletov (@huseletov) 's Twitter Profile Photo

The workplace won’t vanish, it will mutate. What used to be one role becomes a network of micro-tasks-- some human, some not. And someone has to manage them.

Stan Huseletov (@huseletov) 's Twitter Profile Photo

One more industry starts getting reshaped: Keplar raised $3.4M to use voice AI for customer research — faster, cheaper, and so real participants forget it’s a bot techcrunch.com/2025/09/17/kle…

Stan Huseletov (@huseletov) 's Twitter Profile Photo

Behavioral anti-patterns quietly drain outcomes: passive viewing in class, dominance overflow in clinics, activity-theater at work. New video analytics now detect them in real time — lifting quiz scores, boosting adherence, and doubling team output. Invisible inefficiency becomes

Stan Huseletov (@huseletov) 's Twitter Profile Photo

Genome editing just jumped from scalpel to bulldozer. Bridge recombinases mean we’re no longer tweaking genes — we’re reorganizing entire genomic neighborhoods. The leap isn’t just medical, it’s civilizational

Stan Huseletov (@huseletov) 's Twitter Profile Photo

Reasoning isn’t about more tokens, it’s about reusable patterns. ETA shows what happens when models stop re-deriving and start remembering. Procedural memory might be the real unlock for efficient, scalable AI reasoning.

Stan Huseletov (@huseletov) 's Twitter Profile Photo

Prompt hacks break. DSPy doesn’t. By systematically optimizing prompts, it takes schema compliance from ~70% to 90%+ — turning JSON generation from headache to reliable pipeline. Not magic, just repeatable optimization. trilogyai.substack.com/p/useful-or-no…

Stan Huseletov (@huseletov) 's Twitter Profile Photo

Tech literacy didn’t stop Anthony Levandowski from falling into religious delusions, so I don’t think the precautions are exaggerated

Stan Huseletov (@huseletov) 's Twitter Profile Photo

Every time AI is directed toward something genuinely useful for humanity instead of just maximizing profit, it warms my heart. deepmind.google/discover/blog/…

Stan Huseletov (@huseletov) 's Twitter Profile Photo

The most interesting part of AI isn’t what it creates, it’s what it exposes about how fragile our definitions of work and value are.

Stan Huseletov (@huseletov) 's Twitter Profile Photo

Every major tech wave starts with hype about new capabilities. The second act is always about who controls the bottlenecks.