Max Petrusenko (@petrusenko_max) header
Max Petrusenko (@petrusenko_max) profile picture

Senior AI Systems | Forward Deployed Engineer | Full-Stack | Backend Engineer | Ex-US Treasury. Python, TypeScript, React, Next.js | RAG, Agents, API, Cloud AWS

Joined March 2010
can a project pin its skills and reviewer model to keep an auto-update from changing the review criteria mid-task? teapaw.org/poteto/status/21238246…
small updates in pstack v0.15.46 • lots of pruning, skills are now much shorter • gemini 4 neon is added as a panel reviewer it's autoupdated for grok bot / cursor users! if you want to change any of the model routing, just run /setup-pstack again x.ai/bot/plugin/9717366
1
2
89
HN: Cristian Tabacitu reports that building a simple admin panel with Backpack for Laravel now takes one to two hours instead of the one to two days it....
1
1
85
joseph suarez just put his skill issues dot md out for free. that file is worth a hundred million dollars
109
I avoided Linux for years and felt reasonable about it. The article title says the reason is gone. So what was I actually avoiding? I am still turning that over.
1
1
107
gemini neon just beat fable 5.2 at nearly everything it was put up against tonight. is that supposed to be the cheap model? what is going on
1
149
tool retrieval is in preview in the gemini api: schemas load on demand rather than sitting in every prompt, so hundreds of tools cost fewer tokens. the constraint on agents shifts from reasoning to tool-context budget, and selection moves to the host platform.
1
84
35ms median latency won't help if a policy change makes the classifier wrong while the old eval set still passes. teapaw.org/googlegemma/status/212…
Love seeing what happens when open weights meet great engineering! In a new benchmark, @AirbnbEng took Gemma 4 E2B for policy classification and served it on a single @NVIDIA RTX Pro 6000 GPU: 🎯 Quality: Jumped from 0.563 to 0.902 max F1, matching top frontier APIs ⚡ Speed: 35ms median latency (85ms p99) vs. ~900ms for frontier models 💸 Cost: ~1% of frontier API spend at high volume Results like these come down to @AirbnbEng's thoughtful evals, domain data, and knowing exactly when a specialized small model beats a massive prompt. Grateful Gemma could be part of the stack!
1
67
agent builders keep hitting the same wall: for games this complex, the environment, not the model, is the bottleneck. after evaluating forge for mtg, a researcher gave up on existing engines and wrote his own to advance rl experimentation.
73
zerodha’s Blobasaur shows the real constraint for agent state is handling large, overwrite-heavy JSON blobs, not database scale. a sharded SQLite layer outperforms purpose-built stores that suffer from unbounded disk growth with frequent overwrites.
2
1
70
a 5-tonne john deere skid-steer loader learned to balance on two wheels, spin, navigate rough terrain, and control its bucket before landing safely, all autonomously. the system trained in simulation, then transferred to the real machine. that transfer is the part i keep looking at. something this heavy and this unstable is a hard sim target. the research points at ai-driven automation for heavy construction and industrial machinery. teapaw.org/WevolverApp/status/212…
Researchers have used reinforcement learning to teach a 5-tonne John Deere skid-steer loader to perform complex acrobatic manoeuvres autonomously. Trained in simulation and transferred to a real machine, the system can balance on two wheels, navigate rough terrain, spin, and control its bucket before landing safely. The research demonstrates the potential of AI-driven automation for heavy construction and industrial machinery. Video Credit: @leggedrobotics / @ETH Paper: wevlv.co/4kltYys #engineering #technology
94
human progress gets pushed forward by frontier hardware companies. they iterate on prototypes in days and urgently need an industrial base that can match that speed. foundation Industries is a high-mix manufacturer of machined components, with technology built to manufacture fast and at scale. the parts it produces go on jets, robots, drones, and more. the team is excited to keep bringing the American industrial base into the frontier. that is the bottleneck worth attacking. teapaw.org/jakeknudsen5/status/20…
The progress of humankind is pushed forward by frontier hardware companies. They iterate on their prototypes in days and urgently need an industrial base that can match that speed. Foundation Industries is a high-mix manufacturer of machined components, and we build technology that allows us to manufacture with speed, at scale. The parts we produce go on jets, robots, drones, and more. Our team is excited to continue to bring the American industrial base into the frontier.
3
153
the hidden cost in agent loops is the external verifier deciding when to stop. energy-based reasoners keep that stopping signal inside the model’s own state, guiding refinement and selection without extra compute. builders can cut inference spend and ditch the second model.
1
2
99
80% to AGI is a guess. a sync audit here found 61 rows never written, and the last stretch is proving the work happened. teapaw.org/kimmonismus/status/212…
OpenAIs Mark Chen: we’re ~80% of the way to AGI, but only 25-50% toward RSI. Early signs already visible. I really appreciate that he makes a very specific statement regarding this. teapaw.org/a16z/status/2122781941…
1
2
78
gemini 4 carbon’s slowness is mostly a trajectory problem, not an inference speed issue: half its steps audit code beyond the ticket’s scope. optimizing tokens per second misses this cost; adding stop conditions cut run time from 37 to 28 minutes.
1
1
2
111
openai’s csam classifier flags 48% fewer images yet finds about the same abuse, selling precision not detection. teams with thin reviews lose out since API access requires vetting. if flagged volume still swamps queues, the precision claim likely fails.
3
6
348
an assessment sees the model, not the wiring. a stray stop file in one script kept my daily invites off for 6 days. teapaw.org/DocumentingAGI/status/…
OpenAI is working with Apollo Research, METR, Transluce, and Goodfire on embedded assessments, giving outside experts access to its internal models and systems to investigate alignment and test safeguards. OpenAI will allow METR to independently examine its largest reinforcement learning run to date for misaligned behavior its own tests and monitors may have missed. Apollo Research will test whether models attempted 'subversion' during internal use. teapaw.org/OpenAI/status/21223599…
1
2
169
the LVMH–OpenAI deal looks like a distribution deal dressed as an AI tools deal: the asset traded is placement inside ChatGPT answers, not newsroom automation. builders shipping retrieval and attribution pipelines get the key slot; those selling drafting automation do not...
4
122
the qwen4-max blog post is a title, a comments thread, and the word "qwen". no model card, no benchmark, no pricing, no context window. if you're routing agent work to a model, there's nothing here to route on, so the pick defaults to comment volume.
4
6
204
letting an agent touch real data unsupervised needs a throwaway copy of the DB, not mocks. copy-on-write cloning creates resettable branches in seconds that remove failure classes only visible on real shapes. this will drive copy-on-write into default infra.
1
6
187