RT @schickling: Here are some of the things I'd like from A CI system (and what I dislike about the current GitHub actions):. - I want to e….

Claude Code Sub agents have arrived. Can be fine-tuned for domains. while I'm literally reading this at the AI Engineering book by @chipro . Creating small, specialized AI judges as part of the eval systems could be performant & cost effective

Back to civilization. Plug & Play. Simpler than a SIM card. Game changing. @Starlink is what real powerful technology looks like.

Factor #7: Contact humans with tool calls. Sometimes it's preferred to have agent invoke human as "tool". If you use FSM to model human input as well, both human and agent are first class citizen in the systems and orchestration is made easy.

Factor #6: .Launch/Pause/Resume with simple APIs. Send event to the FSM to transition to paused state, wait for next instruction (event).

@statelyai Factor #5: Unify execution state and business state. "Execution state" like current step, status or "Business state" could be nicely structured with hierarchy of FSMs. The state, probably mirrored onto shared datastore, could be easily consumed as future context windows

@statelyai You might have read the classic:. The Log: What every software engineer should know about real-time data's unifying abstraction. Agents execute on events & produce events. Agents could look at the single eventlog, or communicate events in p2p fashion.

Just read up on 12 factor Agents.realize framework like @statelyai might be best to implement many practices. Agents are new, Finite state machine isn't. You want not only intelligence to figure things out, .but also guardrails for deterministic execution.

Language Modeling Is Compression. The Era of Experience.

Terrance Tao --- Physics/Maths Model is Compression of the Universe. Dark Matter model has 14 parameters & explains petabytes of data. Google DeepMind -- LLM is compression. Also this is the era of experience where agent learns from experience. Very inspiring for me to ponder

Revisiting the HK office of my first banking job after 11 years today --- now a coworking spot where I'm building decentralized software. I recall @tkstanczak once mentioned to me he wanted to turn those banking offices into cypherpunk hubs. It's happening. This Leadership🫡.

Nobody wants to fund apps in crypto if reaching "PMF" means apps need to cater for a niche & speculative audience. Nobody wants to build apps in crypto when everyone else wins by being OG or insider and the only value of your app is a showcase link of infra to pump the token.

RT @dDevKit: >Building software for “local” contexts is sometimes easier than building software for world-wide use. > AI is a useful compl….

@VitalikButerin Cant believe it is possible for Socratica to explain things so concisely.

@VitalikButerin Great ROI understanding the intuitions

@VitalikButerin has an excellent premier on finite field for whoever interested.

If you're Dev in crypto, this is a piece of maths you want to understand. Took me a year to read the book. Have to admire Galois's productivity in the sense he invented all this before getting killed in a duel at 20

Best lectures on physics informed ML by @eigensteve .

Agents are useless.