CTO @ Archdesk · ex-hacker · author of Joule Wars (useful intelligence per joule) & Proof-Adjusted Autonomy (PAA) · physics × economics × execution · essays: piszczek.pl/bloghttps://bsky.app/profile/michal-piszczek.bsky.socialhttps://bsky.app/profile/michal-piszczek.bsky.social/post/3mve4ddd35k2yAn AI can improve its own tools without changing its model weights. Does that count as RSI? Sometimes. Does it mean we've built humanity's last necessary invention? That's a much bigger claim. I wrote up the distinction: https://piszczek.pl/blog/rsi-ai-recursive-self-improvement12 Sep 2026 22:17 +0000at://did:plc:errnddtp4wwjm2q56ef37cah/app.bsky.feed.post/3mve4ddd35k2yhttps://bsky.app/profile/michal-piszczek.bsky.social/post/3mtttqono5c25My handle here is a domain I control. Not vanity: provable origin is turning into the scarce input. New essay on where value goes once execution costs nothing. Slop, properly defined, is content nobody is exposed on. https://piszczek.pl/blog/scarcity-never-dies-it-climbs #AI #writing24 Aug 2026 17:36 +0000at://did:plc:errnddtp4wwjm2q56ef37cah/app.bsky.feed.post/3mtttqono5c25https://bsky.app/profile/michal-piszczek.bsky.social/post/3mtcrtzqcfc2zQwen3.8 27B on one 24 GB GPU: full 256K context, really filled, 50 tok/s in production, 2.81x from MTP. Odd finding: a more precise drafter made the system slower. Notes: https://piszczek.pl/blog/qwen38-27b-256k-50-tps-24gb-gpu HF: https://huggingface.co/cdiamond/Qwen3.8-27B-iMatrix-NVFP4-MTP-GGUF17 Aug 2026 22:47 +0000at://did:plc:errnddtp4wwjm2q56ef37cah/app.bsky.feed.post/3mtcrtzqcfc2zhttps://bsky.app/profile/michal-piszczek.bsky.social/post/3msti3doxfk2eI fit Muse Glimmer 30B, vision, DFlash and a real 256K context onto one 24 GB GPU. Code: 17.98 to 84.64 tok/s. Mixed agent work: 38.34. Full KV: 21.56. With speculative decoding, tok/s is partly a predictability benchmark. https://piszczek.pl/blog/dflash-changes-what-tokens-per-second-means11 Aug 2026 20:42 +0000at://did:plc:errnddtp4wwjm2q56ef37cah/app.bsky.feed.post/3msti3doxfk2ehttps://bsky.app/profile/michal-piszczek.bsky.social/post/3ms3v2pgcok2mThe 12 AI economics calculators now run inside Claude and Cursor as an MCP server: npx -y @michalpiszczek/ai-economics-mcp — ask "what does an agent-hour cost?" and it computes instead of guessing. Token cost, LLM energy, PAA, revocation exposure. No key. piszczek.pl/tools02 Aug 2026 11:30 +0000at://did:plc:errnddtp4wwjm2q56ef37cah/app.bsky.feed.post/3ms3v2pgcok2mhttps://bsky.app/profile/michal-piszczek.bsky.social/post/3mrzww6oymk2zHow much does AI actually cost? 12 free calculators — token cost (x177 spread for the same workload), agent-hour, LLM energy, revocation exposure, Proof-Adjusted Autonomy. No sign-up, JSON API + MCP server for agents. https://piszczek.pl/blog/how-much-does-ai-actually-cost01 Aug 2026 16:58 +0000at://did:plc:errnddtp4wwjm2q56ef37cah/app.bsky.feed.post/3mrzww6oymk2zhttps://bsky.app/profile/michal-piszczek.bsky.social/post/3mrxhuqsqdk2tYour token is valid. And it should already be dead. DPoP fixed theft. Stale authority is a different axis — revocation is a propagation race across caches, workers and gateways. Essay: https://piszczek.pl/blog/token-revocation-is-not-an-endpoint Video: https://youtube.com/shorts/rF7HCQHZEqE31 Jul 2026 17:24 +0000at://did:plc:errnddtp4wwjm2q56ef37cah/app.bsky.feed.post/3mrxhuqsqdk2thttps://bsky.app/profile/michal-piszczek.bsky.social/post/3mrjuvzvuqs2cModels get cheaper and more interchangeable every quarter. The harness — the layer holding your context, codebase and workflows — does not. Swap the model, users barely notice. Swap the harness, they lose everything. piszczek.pl/agent-harness26 Jul 2026 07:40 +0000at://did:plc:errnddtp4wwjm2q56ef37cah/app.bsky.feed.post/3mrjuvzvuqs2chttps://bsky.app/profile/michal-piszczek.bsky.social/post/3mr6vncjtw22lAutonomy is not what your agent does. It's what it can prove. A "90% autonomous" agent with 80% evidence, 95% validation and 90% on-time delivery is a 61.6% agent. The metric: https://piszczek.pl/proof-adjusted-autonomy21 Jul 2026 22:54 +0000at://did:plc:errnddtp4wwjm2q56ef37cah/app.bsky.feed.post/3mr6vncjtw22lhttps://bsky.app/profile/michal-piszczek.bsky.social/post/3mr6rak6c3s2iThe model wars are over. The joule wars have begun. Capability is commoditizing — the next axis of AI competition is useful intelligence per joule. The full framework: piszczek.pl/joule-wars Posting from a domain-verified handle now. ⚡ https://piszczek.pl/joule-wars21 Jul 2026 21:35 +0000at://did:plc:errnddtp4wwjm2q56ef37cah/app.bsky.feed.post/3mr6rak6c3s2i