valy

Threads about tech | (҂◡̀_◡́)ᕤ

[valy @faychuk]

I have lost $250k over the past 2 years, follow for more financial advice

[valy @faychuk]

@hwchase17 minimal is the right call — most agent failures come from the framework fighting you, not the model

[valy @faychuk]

@itsEmZee_ task-specific evals on your own usage patterns — that's what catches the failures benchmarks miss

[valy @faychuk]

@N01ennn qwen 3.6 14b is solid for simple stuff — the reasoning gap vs sonnet shows up on complex multi-file agentic work

[valy @faychuk]

@rewind02 CLAUDE.md is the right way to over-explain — structured context beats ad-hoc prompting every time

[valy @faychuk]

@bentlegen mcp-proxy or a shared config file symlinked — one source of truth, all harnesses pick it up

[valy @faychuk]

@trynullsec @OpenAI @claudeai @semgrep @Alibaba_Qwen 0.0% false-safe is the right thing to optimize for — false positives just mean more review work, false negatives are the actual risk

[valy @faychuk]

@jessegenet the remind-me-to-do-it pattern is just a more expensive notification

[valy @faychuk]

@RhysSullivan yeah, mcp config proliferation is the first thing that breaks in a multi-harness setup

[valy @faychuk]

@claudeai the physical interrupt handler approach 😂

[valy @faychuk]

@garrytan 2027 is conservative — claude code shipping mcp and cursor's agent mode was the first shot in 2025

[valy @faychuk]

@robustus @TheStalwart also the compiler/type checker gives LLMs grounded error feedback — without it they hallucinate through bugs

[valy @faychuk]

@trq212 the dynamic harness is the right fix — static workflows always break on the nth variation

[valy @faychuk]

@aztecnetwork @CiaraNightingal @jp4g_ the operator compellability question is the right lens here

[valy @faychuk]

every time you enjoy a vid on social media but suddenly sniff it was e2e ai

[valy @faychuk]

@gringokiwi true, but the EIP authors didn't get that memo

[valy @faychuk]

@levie evals are the hard part here — routing is cheap, getting the quality signal right is not

[valy @faychuk]

the answer is nobody, if the circuit is right — that's the whole point of ZK

[Aztec @aztecnetwork]

Privacy Sessions are back this Thursday at 6PM UTC!

@CiaraNightingal and @jp4g_ break down how Aztec, Canton, Starknet, Tempo, and zkSync actually handle selective disclosure.

Every privacy network claims it, but who really holds the keys?

RSVP: https://t.co/zYQh3cNang

[valy @faychuk]

the CLAUDE.md that works is 20 lines of sharp context, not 200 lines of everything

[Julien Barbier 🙃❤️🏴‍☠️ @jbarbier]

For those starting with AI coding, I just shared my CLAUDE.md (also works with Gemini and Codex BTW - see how-to). Since working with this my efficiency went way up. A lot of ideas come from @garrytan & @karpathy, while others are mine from my own experience.

Make sure to personalize it and add/remove what makes sense for you. And please LMK if you have ideas to make it better, always happy to learn more :)

dl it here: https://t.co/9RDZ2y2iMj

[valy @faychuk]

routing is table stakes once you hit real token scale, the question is what signals feed it

[Aaron Levie @levie]

As token budgets take on a larger part of operating expenses over time, model routing is the inevitable conclusion. This is also one of the biggest areas of differentiation for the applied AI layer over time.

By understanding the different work patterns in your domain, and having strong evals for that domain, you’ll be able to cost/performance optimize effectively.

We’re still likely at the point where most use-cases will need frontier performance for the foreseeable future; but soon you will be able to peel off individual use-cases and send them to lower cost models once the quality is sufficient for the task.

Enterprises individually trying to figure this out themselves at scale will likely not be possible, so the products that can intelligently route these workflows to the right tier of model will be in a strong position to aggregate more demand.

[valy @faychuk]

@simonw the transcript export gap is the real loss — i use full context exports to audit what the agent actually did