Today @ConcordanceAI is open sourcing our modified inference engine for token-level interventions, alongside an injection playground and supporting research on steering LLMs.

We find that token injection is indeed a viable steering mechanism, in certain cases marginally outperforming user prompting in effectiveness that we explore in our research. The Concordance engine goes beyond simpler token injection methods by stacking several primitives together including:

Force Tokens
Adjust Logits
Backtrack

This allows for multi step injections, find & replace, and structured outputs, occurring at a various points in the generation stream, not just the beginning.

Here is an example of injecting a "double check" steering sentence: research.concordance.co/share/request/…

you're writing your smart contracts wrong and leaving your protocol open to vulnerabilities

introducing FREI-PI: a new defi smart contract security pattern, the logical successor to Checks-Effects-Interactions

give a read of my blog post about it:

nascent.xyz/idea/youre-wri… why has Uniswap never been hacked?

why was dYdX's SoloMargin never hacked?

why was Euler hacked for $200m?

why was Compound's cETH market unusable for a couple weeks?

what is the difference maker?

solana claims to be able to handle more transactions thanks to their parallel execution

how big of a speed up could the evm get if it did something similar?

i built a parallel evm execution engine to find out

/thread first, some background. the evm is “Virtual Machine”, or software that simulates hardware, that standardizes and simplifies what is possible to do on the Ethereum blockchain