Joe Fox

I research entrepreneurship, innovation, and AI. Following calling w @Alexakaye3 & thanking God for it. Like != endorsement

[Joe Fox @JosephDFox]

Χριστὸς ἀνέστη

[Department of Government Efficiency @DOGE]

Great collab: @USGSA x @VP, @AFergusonFTC

[Joe Fox @JosephDFox]

Apparently March is agent finance stack announcement month

[Joe Fox @JosephDFox]

A lot of AI content can be categorized as shiny coins or dull $100 bills often taking the same amount of time to consume.

[Joe Fox @JosephDFox]

Agentic organizations & companies as well as governance & orchestration approaches have me thinking more about org culture. Lots of individual approaches on agent persona/values, surprised there isn't a pip repo to distill "culture" from one group of agents to another. Could easily see the agentic persona marketplace moving into these areas.

[Joe Fox @JosephDFox]

Interesting on persona dropout technique: https://t.co/pzwUnusPgP

[Shirley Wu @ShirleyYXWu]

We share two blogs outside of the HumanLM paper: https://t.co/K0pOqeMPal

Is Synthetic Data Good Enough to Train User Simulators? — by me and @ArpandeepKhatua

Persona Dropout Makes Robust User Simulators @Es2C003

+ Code is ready here! https://t.co/Iw4eRh63gX

[Joe Fox @JosephDFox]

Should skill md files have "nutritional facts" or outline of risks?

[Stephen Wolfram @stephen_wolfram]

Making Wolfram tech available as a Foundation Tool for LLM foundation models...
https://t.co/TbJ4RIjADG

[Joe Fox @JosephDFox]

Have you seen any physical /stop buttons for OpenClaw yet?

[Robert Scoble @Scobleizer]

What was I talking about yesterday?

OpenAI can put @openclaw on a small device.

I could see buying several and leaving one in each of our cars.

[Joe Fox @JosephDFox]

What websites do OpenClaw agents visit for vacation?

[Joe Fox @JosephDFox]

Cool approach to deal with fragmented PII snippets across context

[Hyunwoo Kim @hyunw_kim]

🚨New paper to level up your 🦞#Clawdbot ?!
Bots are now posting your sensitive info in real time. But privacy research is a desert with no data to train better models. That's about to change

Enter 🏝️Privasis, the oasis where you can train strong privacy-forward AI with scale✨

[Mixedbread @mixedbreadai]

The multi vector embeddings are created using our SOTA retrieval model mxbai-wholembed. It supports text, audio, video on over 300 languages. A key innovation is the dynamic vector allocation, which lets the model dynamically decide the amounts of vectors it needs to represent information. For example, a simple cat image may output a few vectors, whereas a complex slide deck may generate thousands of vectors. We wrote a custom inference engine to serve mxbai-wholembed with low latency.

[Joe Fox @JosephDFox]

@joshgans @RemindMe_OfThis in one week.

[Joe Fox @JosephDFox]

@m_finzi @ShikaiQiu @yidingjiang @Pavel_Izmailov @zicokolter @andrewgwils @RemindMe_OfThis to read in 1 week.

[Adam Karvonen @a_karvonen]

New paper!

A compromised inference server can leak model weights by hiding them in normal-looking responses🥷

But LLM inference is nearly deterministic, so we can verify outputs with a trusted server & detect this (and inference bugs)

This slows data exfiltration by >200x

🧵

[Justin Gordon @justingordon212]

I want to update our list of resources for founders raising capital

What should we include?