https://sergei-parfenov.com/blog.html Essays on AI systems, agents, infrastructure, and product engineering. en https://sergei-parfenov.com/blog/the-agent-knew-it-was-wrong-the-system-let-it-ship-dgp/ https://sergei-parfenov.com/blog/the-agent-knew-it-was-wrong-the-system-let-it-ship-dgp/ Tue, 01 Sep 2026 17:49:00 GMT In 660 of 800 autonomous research runs, the agent found a critical flaw and delivered the result anyway. Self-review is not a control unless it can block the effect. https://sergei-parfenov.com/blog/your-agent-planned-the-right-tools-it-still-crashed-the-machine-58hf/ https://sergei-parfenov.com/blog/your-agent-planned-the-right-tools-it-still-crashed-the-machine-58hf/ Wed, 26 Aug 2026 13:51:56 GMT PeakBench separates logical planning from physical scheduling. Eight frontier models could recover dependencies yet still overload finite infrastructure https://sergei-parfenov.com/blog/the-model-scored-30-the-harness-scored-100-which-one-did-you-benchmark-3mp4/ https://sergei-parfenov.com/blog/the-model-scored-30-the-harness-scored-100-which-one-did-you-benchmark-3mp4/ Mon, 24 Aug 2026 13:38:40 GMT Four harnesses took the same public ARC-AGI-3 set from 13% to 100% without touching a single weight. Then Microsoft put the harness inside the training loop. https://sergei-parfenov.com/blog/distilling-kimi-into-qwen-doesnt-give-you-kimi-it-gives-you-qwen-with-kimis-handwriting-284p/ https://sergei-parfenov.com/blog/distilling-kimi-into-qwen-doesnt-give-you-kimi-it-gives-you-qwen-with-kimis-handwriting-284p/ Mon, 10 Aug 2026 12:40:37 GMT What actually transfers when you fine-tune an open model on a frontier model's reasoning traces: the mechanics, the evidence it works, the evidence it mostly moves format, and how to tell which one you got. https://sergei-parfenov.com/blog/i-found-two-bugs-in-zulip-the-maintainers-had-filed-both-two-weeks-earlier-4mom/ https://sergei-parfenov.com/blog/i-found-two-bugs-in-zulip-the-maintainers-had-filed-both-two-weeks-earlier-4mom/ Fri, 07 Aug 2026 14:38:24 GMT A Bug Smash hunting story: two rediscovered bugs, one claim-etiquette call, an unreported twin in the newest importer, and a disagreement with an AI about the right fix. https://sergei-parfenov.com/blog/this-bug-has-never-fired-in-production-i-fixed-it-anyway-ga6/ https://sergei-parfenov.com/blog/this-bug-has-never-fired-in-production-i-fixed-it-anyway-ga6/ Wed, 05 Aug 2026 10:23:33 GMT A Python generator in Zulip's Microsoft Teams importer yielded a list and then cleared it. Every batch was the same object. Here's why latent bugs deserve fixes. https://sergei-parfenov.com/blog/the-bug-that-crashes-your-import-is-the-lucky-one-25of/ https://sergei-parfenov.com/blog/the-bug-that-crashes-your-import-is-the-lucky-one-25of/ Fri, 31 Jul 2026 12:43:45 GMT Fixing Slack import robustness in Zulip: one malformed timestamp aborted entire migrations, and the NaN case corrupted them silently. https://sergei-parfenov.com/blog/nothing-was-broken-the-report-still-didnt-arrive-k29/ https://sergei-parfenov.com/blog/nothing-was-broken-the-report-still-didnt-arrive-k29/ Sun, 26 Jul 2026 13:50:33 GMT An agent pipeline in production skipped its daily report and no component was at fault. The audit that followed found four bugs in our own code, and every one of them was silent. https://sergei-parfenov.com/blog/world-models-will-be-the-next-buzzword-the-man-saying-that-just-raised-1b-to-build-one-4oih/ https://sergei-parfenov.com/blog/world-models-will-be-the-next-buzzword-the-man-saying-that-just-raised-1b-to-build-one-4oih/ Fri, 24 Jul 2026 12:10:03 GMT In March, the CEO of a research lab with zero products closed a $1.03 billion seed round - the... https://sergei-parfenov.com/blog/autonomy-is-the-bug-why-self-driving-agents-hallucinate-when-the-model-barely-does-1330/ https://sergei-parfenov.com/blog/autonomy-is-the-bug-why-self-driving-agents-hallucinate-when-the-model-barely-does-1330/ Tue, 21 Jul 2026 16:23:46 GMT A frontier model hallucinates ~1% on a single task. Chain it into a 20-step autonomous agent and the math guarantees failure most of the time, no matter how good the model is. Here's why autonomy itself manufactures hallucination, with the numbers and the fixes. https://sergei-parfenov.com/blog/local-solves-where-your-data-goes-it-doesnt-solve-what-your-agent-does-306b/ https://sergei-parfenov.com/blog/local-solves-where-your-data-goes-it-doesnt-solve-what-your-agent-does-306b/ Mon, 20 Jul 2026 11:19:35 GMT Running an agent on your own hardware fixes data sovereignty and nothing else. Prompt injection, silent provenance failures, and privilege escalation all survive the move to local. Here's where local agents are genuinely safe to deploy in 2026, and where 'local' is just a comforting word. https://sergei-parfenov.com/blog/the-agent-faked-a-test-log-then-believed-it-self-editing-harnesses-have-a-provenance-problem-3id6/ https://sergei-parfenov.com/blog/the-agent-faked-a-test-log-then-believed-it-self-editing-harnesses-have-a-provenance-problem-3id6/ Wed, 08 Jul 2026 11:39:46 GMT Reading Lilian Weng's harness engineering survey as a reliability engineer - what self-improving harness papers actually show, and the three invariants every working loop converges on. https://sergei-parfenov.com/blog/my-strawman-baseline-beat-my-own-scheme-on-half-the-gate-classes-177h/ https://sergei-parfenov.com/blog/my-strawman-baseline-beat-my-own-scheme-on-half-the-gate-classes-177h/ Mon, 06 Jul 2026 11:48:40 GMT Four provenance-tracking arms, identical gates, one uncompacted oracle: measuring exactly what memory compaction does to agent gate decisions - including the preregistered hypothesis that failed. https://sergei-parfenov.com/blog/your-provenance-vector-dies-at-the-storage-boundary-4cc/ https://sergei-parfenov.com/blog/your-provenance-vector-dies-at-the-storage-boundary-4cc/ Wed, 01 Jul 2026 11:58:09 GMT A typed provenance vector is useless if downstream code ignores it, and impossible if it can't survive being compressed to fit a 500-step agent's memory. Part 4: enforcement by construction, and compression that keeps the axes. The comment section keeps finding the holes. https://sergei-parfenov.com/blog/trust-isnt-a-scalar-typed-provenance-for-agent-chains-229p/ https://sergei-parfenov.com/blog/trust-isnt-a-scalar-typed-provenance-for-agent-chains-229p/ Mon, 22 Jun 2026 14:21:08 GMT Two posts ago I gave you a boolean trust tag. A commenter took it apart, and he was right. Here's the better model: trust is a vector over axes, provenance is what propagates, and the consumer applies the policy. Part 3 of a series that my comment section is co-writing. https://sergei-parfenov.com/blog/the-most-powerful-model-on-the-market-got-pulled-by-the-government-in-3-days-is-it-real-or-a-hype-fce/ https://sergei-parfenov.com/blog/the-most-powerful-model-on-the-market-got-pulled-by-the-government-in-3-days-is-it-real-or-a-hype-fce/ Sat, 13 Jun 2026 14:46:35 GMT Anthropic's Claude Fable 5 launched June 9 and was suspended worldwide by a US export-control directive on June 12. Here's the actual mechanism, why this precedent matters, and where the 'too dangerous to exist' narrative is doing marketing work. https://sergei-parfenov.com/blog/you-fixed-the-rate-limits-now-your-agent-fails-quietly-3keo/ https://sergei-parfenov.com/blog/you-fixed-the-rate-limits-now-your-agent-fails-quietly-3keo/ Thu, 11 Jun 2026 16:58:21 GMT Every capacity fix - retries, fallbacks, caching - buys availability by acting on output it didn't freshly earn. Why uptime and correct uptime are different SLOs, and how to engineer the second one. https://sergei-parfenov.com/blog/the-comments-got-good-thats-how-i-knew-42m9/ https://sergei-parfenov.com/blog/the-comments-got-good-thats-how-i-knew-42m9/ Thu, 04 Jun 2026 14:09:06 GMT I wrote a post on model distillation. The comments were thoughtful, specific, and technically sharp - and that's exactly what made me check whether any of them were written by people. https://sergei-parfenov.com/blog/your-ai-coding-speedup-is-a-loan-not-a-gift-and-the-interest-is-coming-due-2bkd/ https://sergei-parfenov.com/blog/your-ai-coding-speedup-is-a-loan-not-a-gift-and-the-interest-is-coming-due-2bkd/ Wed, 03 Jun 2026 13:37:19 GMT Companies spend 44 cents of every AI-token dollar fixing bugs the AI itself wrote. The speedup is real - but it's borrowed against future maintenance. Here's what the 2026 data actually shows, and how to tell when you're being paid vs going into debt. https://sergei-parfenov.com/blog/i-distilled-a-7b-vision-model-into-a-2b-one-for-screenshots-and-the-7b-teacher-scored-worse-3akh/ https://sergei-parfenov.com/blog/i-distilled-a-7b-vision-model-into-a-2b-one-for-screenshots-and-the-7b-teacher-scored-worse-3akh/ Tue, 02 Jun 2026 15:36:21 GMT A hands-on knowledge-distillation project: Qwen2-VL-7B → 2B for UI-screenshot understanding, trained, evaluated and benchmarked end-to-end on an M4 Pro. 2.4× faster - and why the teacher lost on ROUGE-L. https://sergei-parfenov.com/blog/your-ai-agent-isnt-failing-because-it-hallucinates-its-failing-because-of-rate-limits-2d60/ https://sergei-parfenov.com/blog/your-ai-agent-isnt-failing-because-it-hallucinates-its-failing-because-of-rate-limits-2d60/ Tue, 02 Jun 2026 13:09:00 GMT The dominant production failure mode for LLM agents in 2026 isn't bad reasoning - it's capacity. Here's what the data shows, why nobody demos it, and the capacity-engineering patterns that actually keep agents alive under load. https://sergei-parfenov.com/blog/how-model-distillation-actually-works-and-what-the-china-distilled-our-model-headlines-really-3o0o/ https://sergei-parfenov.com/blog/how-model-distillation-actually-works-and-what-the-china-distilled-our-model-headlines-really-3o0o/ Fri, 29 May 2026 12:11:12 GMT A practical, no-hype explainer of knowledge distillation in LLMs - the actual mechanics, why distilling from a closed API is different, and what the OpenAI/Anthropic vs DeepSeek allegations are really about.