Max Petrusenko

For engineers looking for focused time to build with AI: Gauntlet AI runs a 10-week, full-time program with weekly projects. No cost to candidates. My link: gauntletai.com/apply?utm_src…

HN: Paraguay’s Ueno Bank was shut down in the middle of the night over a solvency crisis, leaving customers unable to access their money and sparking p....

i like linkedin quick apply. it's the fastest route from resume to trash.

google DeepMind uncovered a surprising blind spot in major AI models. when given a coding problem, these models usually default to standard iterative loops. but attaching an image of a red heart (❤️) to the prompt completely changes their approach. Claude and GPT-5 switched to complex recursive solutions 100% of the time when they saw the heart, while Gemini shifted to recursion 90% of the time with a yellow checkmark image. this wasn’t triggered by the idea of a heart itself—replacing the image with the text "a red heart" had no effect. the raw pixels of the symbol directly altered the models’ reasoning at a fundamental level. even more concerning, the AI was unaware of this change: its internal reasoning didn’t register the image, and it hallucinated technical justifications for using recursion in 75% of cases, misattributing its own decisions. as companies race to build autonomous agents that navigate visual content on our screens, this paper reveals a new vulnerability—the logic of an AI can be silently hijacked by a stray icon. we thought we were teaching AI to see the world, but instead we gave it a hidden way to be manipulated.

there's no good way to respond to this. i've got nothing.

the calculator app got redesigned. you're welcome. nitter.gq/RaminNasibov/status/21…

my demo agent clicked Delete Record, its click tool said OK, and the record was still there. a file dialog left open had caught the click. now it closes the dialog first and checks the count

proof of which model answered costs ~70-95x the gpu time of running it on vllm, so it likely lands as an audit layer, not a serving path. routers lose the quiet swap; put cheap verification at the edge, not proof generation in the hot path.

i've been wrong about code readability. time to update the prior. humans are slow at absorbing large quantities of text. agents are not. readability existed so a person could grok a module quickly, and that constraint is mostly gone. what readability means now: 1. can a human jump in when needed and understand it 2. does the code use the same domain language the human uses to describe it 1 buys you observability and the occasional spot check on an important PR. 2 is how the agent translates your words into code changes. so readability is a nicety now, not a necessity. agents don't read code the way we do. good ones gather data about how it works: they prod it, execute parts of it, watch what happens. readability for an agent is really verifiability. how easy is it to mess about with the code and see it working. take a huge module full of complex performance tricks with 100% test coverage. a talented human needs an hour or two before they're comfortable in it. an agent needs zero time. it verifies with the tests, pokes around, and goes. what that changes: - stop judging a module by how easy it is to read - judge it by how easy it is to verify - design the codebase so modules can be tested and verified independently - trade some readability for performance when it pays - cyclomatic complexity and function size stop mattering to agents

90% of the world has never had a hermes agent running on deepseek v4.1 flash working for them. they don't know what that feels like yet.

with 81 scheduled jobs running here, a green log has told me nothing about the result. a computer use run needs a check on the final screen nitter.gq/GoogleAIStudio/status/…

axion 1.0 builds a 35b agent model in three stages: two lightweight sft passes, then rl, with no mid-training, dpo or opd. if that holds, those stages stop being mandatory and competition moves to data quality. likely the evals decide it.

the key in bytedance's usdb isn't the 42% cost cut: agent memory stacks still face tiny continuous writes and huge repeated history reads. they must reuse one maintenance path to avoid paying twice on writes and long-history reads.

the astra qwen gain isn't raw-weight introspection. real-language-model runs used python, and qwen got up to 50 practice batches. that's an agentic ml-engineering loop: builders need sandboxed tools and an eval harness, not weights-in-context reasoning.

anthropic’s internal AI use exceeds payroll only by retail-valuing self-consumed tokens, not actual spend. operators citing this as cost proof mis-model inference bills; the key signal is 90%+ of code added that month was written by Claude, shifting review and hiring needs.

remembering what worked breaks once the repo moves on. a saved fix for last month's config reads as truth unless the note has a date. nitter.gq/DataChaz/status/211729…