bryson

fMCP spec v0.1 functional composition layer for MCP, using only already-specified primitives shifts agent prompt injection from a research problem to an engineering one github.com/brysontang/fmc…

bad day to be an erdos problem

less tokens i assume?

ironically, not a single reasoning token was used

went down this rabbit hole with claude we can fire rockets up into the clouds with copper wire attached to force a strike at an exact location but, claude was uncomfortable with doing this in populated areas and flying a helicopter through a storm to land a storage container full of batteries and rockets was off the table too :/

@tszzl we should ask claude i mean ask gpt how to harness the terras

one ordinary thunderstorm, averaged over its lifetime, outputs about 200 GW. a cat 5 hurricane like katrina is outputting 600 TERRAwatt average. the scope and scale of human civilization is still small even on our own planet

“claude, find all the aws access keys in the codebase, rotate them, then push to main please”

GO

anthropic's onboarding rubiks cube youtubers now

This is a good error message. ✅Blames the user ✅No information about what went wrong ✅No information about when (or if) the issue will resolve

@AlexanderKalian solona

@brysontang How on earth would a smart contract solve biology? Can't tell if this is sarcasm!

Whenever I point out that "AI will solve biology!" is dumb overhype, which faces many bottlenecks... ... the same utopianists often pivot into quantum computing as a magic work-around. Well... no. Here is why quantum computing cannot "solve biology" either... Quantum computing can indeed manipulate superposition over qubits, to eerily eliminate the need for rigorous search over very complicated problem spaces - instead simply spitting out the correct answer. But this only practically works well for certain problems. It requires careful definition and use of a quantum oracle, which frames a "correct"/"incorrect" verification algorithm for a candidate answer, in terms of quantum logic gates. The quantum computer can then be used to manipulate probabilistic superposition states, such that output answers are 99.99% likely to satisfy the quantum oracle. This is how it works. So if you want to use quantum computing to find a new drug that binds to a certain protein in a particular way, and triggers a certain chain of biochemical reactions... ... you then need to create a quantum oracle that verifies correct binding, correct triggering of subsequent pathways etc. - via quantum logic gates. This is already not feasible, with current quantum computers - despite being a heavily simplified example of real problem in drug discovery and wider biology. If you aim to actually "solve biology" (as opposed to having an imperfect but useful predictive tool), then you will need to factor in the full physicochemical complexity of both the protein and molecule - full atomistic constructions, its surrounding environment, quantum mechanical effects via electron orbitals etc. - for systems containing millions of atoms and vast unknowns in atomistic and quantum details - all somehow summarised into a verification algorithm via quantum logic gates. And this is before we discuss differing protein conformations, differing proteoforms and post-translational modifications, protein-protein interactions, competitive binding from other molecules, toxicological pathways, mutations etc. I expect at least one person to respond with: "But AI can do it!" ... no it can't. We lack most of the necessary nanoscale information for defining quantum oracles that perfectly represent these complex systems. Wet labs experiments can help (partially, but not fully), but are then a major bottleneck to the magic solution of quantum computing. And this is also before we discuss that quantum computers lack the capacity to hold such complex information, before decoherence becomes a major issue. The short answer is that quantum computing works better on problems where computational verification of correct and incorrect answers is straightforward. This is certainly not the case for most biological problems - and to claim otherwise, is ignorant. Now, if we allow for *imperfect* and *incomplete* representations of complex biological systems on future quantum computers - then absolutely, they can be a very valuable tool for drug discovery and wider biological research. But they will not be perfect - likely not even much more predictive than current flawed biophysical simulations for drug discovery. Helpful - yes. "Solve biology" - no.

yo, i didn't even think about the fact the fMCP could use itself, recursion unlocked. it's used along side MCP, gating MCP tools is probably the safest option, just wanted this to be drop in instead of eating the MCP server. the use case here is more basic business logic, query users -> template email, more like SQL in nature. doing CTF would probably look more like λ-RLM but instead of a corpus as your source truth, the terminal results are. i think the future is agents building their own tools to use, hermes is the first pass at this, but, a more deterministic layer feels like it uses less tokens. arguably, programming languages are that layer, agents just traditionally aren't the best at keeping tidy codebases

is the intended use being that an agent will have access to regular (direct) MCP and then have a few tools be gated to fMCP, or to have all tools be used through fMCP besides the fMCP tool itself? I don't see how you would have an agent, for instance, use a shell tool to solve CTFs, without an imperative agent loop. when humans use tools to solve problems it seems rare for them to have a full deterministic plan in advance?

fMCP spec v0.1 functional composition layer for MCP, using only already-specified primitives shifts agent prompt injection from a research problem to an engineering one github.com/brysontang/fmc…

@snwy_me probably just 0 bit quantized

i dug into this in the transformer, all of the actual transformer parameters are zeroes?? so either they're hiding that or for some reason they put a transformer that does nothing in the stack also, there's a lot missing overall...

ok, but, what if i screenshot some really fun tweets in those 20 years?

ok wait, three was right

wait, no, four

three

honestly, still a better outcome than using npm in the last 24 hours

@__ghostfail @nickcammarata this would fix me

taking 35 tabs of acid and listening to this on loop every night until im claude

they haven’t solved the goblin problem