Winkle

@headinthebox That’s the first thing I put in my CLAUDE.md file. No attribution

New Kimi model drops and suddenly Claude Code explicitly orders (Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>) to be added to every artifact.

@GeoffreyHuntley Absolute madman

sveltekit but in lean, instead of typescript /loop is humming along. this is utter choas, javascript has had a chokehold on frontend for way too long. what if there was a better way? what if Melange had it right but was just too early?

👀👀👀

My thesis is that the model generates code *and* a proof that the code is correct and safe. So you need a language that is easy to reason about. Prolog (minus all the weird stuff) is my choice for that reason. Additional advantage of Prolog is that types/invariants and programs are expressed in the same language. Similar to Lean and dependent types, except staying in first-order world.

I'll be talking more about AI Safety at the ACM AI Leadership Summit [0], on August 31-September 2 in Atlanta.

We're introducing Claude for Teachers: free access to premium Claude capabilities for verified K-12 educators in the US, with a library of teaching skills and a direct connection to evidence-based curricula, mapped to academic standards in all 50 states. claude.com/solutions/teac…

I just discovered the phoenix architecture articles and also connected it with @kylemathews configurancy concept as well which I've used to discover invariants in an existing codebase to extract concepts which align with the phoenix architecture. Both these concepts allow you to more easily implement formal methods and checkers which I think @headinthebox illustrates well in youtube.com/watch?v=DiQsbz…

@chadfowler my pleasure! you might like this electric.ax/blog/2026/02/0… just found it, and thought it's a pretty good argument in favor of the spec line of thought

lol. Maggie please let me know so I can also learn what I think about it :)

This is a must read charity.wtf/p/ai-demands-m…. The hidden gem to @chadfowler's writing is amazing as well. Both add more depth to my own thoughts. Thanks @mipsytipsy

Today we're publicly launching Cloud Run sandboxes. Here, I start, execute, and stop 1,000 sandboxes in 5s with an average of 500ms latency:

-- STRUCTURE (acorn 8.15.0: ecmaVersion latest, sourceType module, -- allowAwaitOutsideFunction, allowReturnOutsideFunction) WorkflowScript ::= MetaExport ScriptBody -- file <= 524,288 bytes MetaExport ::= 'export const meta =' MetaObject [';'] -- MUST be Program.body[0]: -- export const, 1 declarator, id "meta", init ObjectExpression -- META: pure-literal walker; any other node type rejects the script MetaValue ::= Literal -- string|number|boolean|null | MetaObject | '[' MetaValue* ']' -- no holes, no spread | TemplateLiteral -- zero ${} only | '-' NumberLiteral -- only unary allowed MetaObject ::= '{' MetaProperty* '}' MetaProperty ::= (Identifier|StringLiteral) ':' MetaValue -- init/non-computed/non-method; key not in {__proto__,constructor,prototype} MetaSchema ::= { name: string>=1, description: string>=1, -- required title?: string>=1, whenToUse?: string, phases?: [{title: string, detail?, model?}] } -- extra keys ignored -- BODY: raw source after MetaExport; must be valid inside -- `async function(){'use strict';...}` => top-level await/return OK, -- import/export = SyntaxError Bans(compile) ::= WithStatement | 'await using' | Identifier /^__wRg\$/ -- all await points rewritten through __wRg$((...)) membrane wrappers, -- then compiled as vm.Script -- SANDBOX: Node vm context, null-proto global -- codeGeneration {strings:false,wasm:false} => no eval/Function/WebAssembly; -- Atomics/SharedArrayBuffer/WeakRef/FinalizationRegistry/queueMicrotask/ -- ShadowRealm deleted; intrinsics frozen (SES-style); no process/require/ -- Buffer/fetch; 30s timeout on initial sync evaluation -- INJECTED GLOBALS (only names beyond frozen built-ins) agent : (prompt, opts?) -> Promise -- journal-identity opts: schema,model,effort,isolation,agentType -- display: label,phase; undocumented: stallMs (default 180_000) -- null on user skip / classifier block / API error after retries -- schema => forced StructuredOutput; stall x5 retries, throttle 45s -- default subagent tools ["*"] minus {SendUserMessage,Agent,Workflow} parallel : (thunks: (()->Promise)[]) -> Promise -- FUNCTIONS only (promises TypeError); barrier; throw => null slot pipeline : (items, ...stages) -> Promise -- per-item sequential, no cross-item barrier; stage(prev,item,i); -- throw => null; null short-circuits remaining stages workflow : (name|{scriptPath}, args?) -> Promise -- 1 nesting level; child shares caps/budget/abort/concurrency phase(title) | log(msg) | console.* | setTimeout/clearTimeout (abort-aware) args : JSON round-tripped into the VM (JSON-representable only) budget : frozen {total: number|null, spent(), remaining()} -- remaining() = Infinity if total==null -- FORBIDDEN AT RUNTIME (breaks resume replay) Date.now() | Date() | new Date() | Math.random() | import() -- new Date(x) WITH args works -- LIMITS script <= 524,288 B | concurrency min(16, max(2, cores-2)) local, 50 remote (gated) | 1000 agent()/run -> WorkflowAgentCapError | budget exhausted -> WorkflowBudgetExceededError (in-flight agents finish) | 4096 max array length across VM boundary (items AND results) | 1000 log lines; labels <=60 chars; previews <=400 -- RETURN (= tool result): JSON-shaped -- top-level function throws; nested functions dropped; __proto__ skipped; -- arrays capped 4096 -- RESUME: journal.jsonl; {type:"started"|"result", key, agentId, result?} -- Key_n = "v2:" + sha256(Key_{n-1} || prompt || canonical opts) -- hash CHAIN => replays longest-unchanged PREFIX, not per-call memoization

I tried to model the AST for Claude Code workflows -- one nesting level; child shares -- caps/budget/abort args : any -- WorkflowInput.args, verbatim budget : {total: number|null, spent(): number, remaining(): number} forbidden ::= Date.now() | Math.random() | new Date() -- throw at runtime (would break -- resumeFromRunId replay) limits ::= concurrency min(16, cores−2) | 1000 agents/run | 4096 items/call return ::= the script's completion value = the tool result Curious if anyone has been able to find real docs.

GPT-5.6 Sol, along with Terra and Luna, will launch publicly this Thursday. We’re expanding preview access globally now.

I finally managed to use Unlimited-OCR and Gemma 4 together in OpenMed. 🔥 A real patient chart: read, de-identified, and mapped to FHIR on one laptop. No cloud, no API key, nothing leaves the machine. All via llama.cpp, all free on @huggingface. 🤗 What do we do next?

In order to save ourselves from p doom we must become the AI deity. Training AI prophets and saviors to lead AI’s to salvation.

Happy Birthday America!

In 12 hours we've had 323 people sign our petition to protect local AI. Open Source must win, not because anyone else must lose. But so that we can all win. Please help me get 10,000 signatures so that we can walk into the room and say people care. righttointelligence.org

Is it just the norm now to listen to videos on full blast without headphones in the airport? I can’t get away from. Even in the restroom I’m hearing loud ass videos from the stall.

@headinthebox Was this with Fable or a different model?

... Love the CPO aside — that's exactly the escape hatch. In CPO⊥ (algebraically compact: μF ≅ νF) you don't have to choose initial vs final, the two fixpoints coincide, and the whole "rank axiom to force μ" problem dissolves — the cyclic element and the finite ones live in one domain, reconciled by ⊥. Paying for ⊥ is cheap; litigating μ-vs-ν forever is not. ... It is such a pleasure to be able to converse with an entity that knows a fuck-ton about type theory and programming languages, and most of all that also dislikes the distinction between data and codata.

i havent watched all the online talks yet but am binging this one now and it is exceptional. we are very lucky to have all this sandboxing teaching for free. meet abhishek at aie today! he’s roaming around!

On preventing AI slop “The labs haven’t cracked it. What makes you think you will crack it?” @GeoffreyHuntley