Christian

Professional snack thief ⊹ building cool stuff at

[Christian @imCBV]

agentic web search should be treated as poisonous by default. blogs, listicles, AI-generated summaries: most of what comes back is either marketing or even straight malicious. here's what I actually do when I research with AI:

[Christian @imCBV]

@chedisagio @GabrieleCarrer @ZohranKMamdani @pallonatefaccia materiale per te

[Christian @imCBV]

@compileandpush Nuke the site, the ones that desperately get in touch are prime material for user interviews 🧠✨

[Christian @imCBV]

slop docs flying around, slop images all over social, open source drowning in slop code. it just shows what we always knew: generative AI is a tool, and how you use it makes all the difference

[Christian @imCBV]

@IntCyberDigest Track ID?

[Computer ♥ Records @ComputerLove_]

Nissan COM COM Concept Van (1985) A prototype van with ‘advanced communications capabilities’ including an early satellite navigation system.

[Christian @imCBV]

here's how i actually handle it: treat everything an agent fetches as hostile, and never let the agent that read it be the one that runs anything. full approach here
https://t.co/wGchLSi1WV

[Christian @imCBV]

earlier this week i said treat anything your AI agent fetches as poisoned. guess what: a dev hid instructions in their own code telling any AI that copies it to wipe your data
https://t.co/Jwz3UTa4Pf

[Christian @imCBV]

@oizo3000 Isn’t that the YouTube dubbing?

[Christian @imCBV]

the first thing I do when starting a new task is to get an agent to research what experts+community are saying, something I did for ages without thinking until I realized how naive it is: four sources agreeing means nothing if they all copy one blog post
https://t.co/SMIb834p7P

[Christian @imCBV]

@kavinsky What’s up?

[Christian @imCBV]

@AxialisSoftware @trq212 Yeah LLMs still struggle with that but I'm mostly using this approach for local throwaway explainers, nothing that ends up in production

[Christian @imCBV]

i find @trq212 approach really interesting: ask Claude/Codex for html output instead of markdown, you get .svg diagrams inline, anchor links, interactive widgets, callouts, the markdown default is a habit from old models that makes no sense at modern context sizes

[Christian @imCBV]

compiling all this into a skill I'll release on my github soon. anything I missed?

[Christian @imCBV]

agentic web search should be treated as poisonous by default. blogs, listicles, AI-generated summaries: most of what comes back is either marketing or even straight malicious. here's what I actually do when I research with AI:

[Christian @imCBV]

finally for code or configs, use a second agent to audit the proposed code without showing it the original blog posts. malicious instructions buried in retrieved pages should never reach the implementer that way

[Josef Chen @josefchen]

Launching our new paper on arXiv: we trained the largest multilingual food model ever built.

4.1M recipes. 7 languages. 1,790 ingredients. 300 dimensions.

All of human cooking compressed into 2 megabytes.