They're entirely untrustworthy for anything like access control, because whoever gets control of a portion of the input tokens effectively controls the output https://
simonwillison.net/2023/Nov/27/pr
ompt-injection-explained/
…
LLMS
-
Prompt Injection Security Vulnerability in AI Access Control Systems
By
–
-
LLM Corporate Use Cases: Data Extraction Over Safety Trade-offs
By
–
I wonder about that: if you're building public-facing chatbots then sure, you want everything as safe as possible – but from what I've seen the most valuable corporate use-case for LLMs is structured data extraction, where any denials are just a huge nuisance
-
LLMs struggle with counting tasks in images
By
–
Given how bad LLMs are at counting everything else – words, letters, etc – I'd be surprised if they can reliably count things in images
-
Personal AI Meetup Marks Inflection Point in AI Technology
By
–
Just hosted the world’s first Personal AI meetup! (where recording is opt out instead of opt in)
— swyx 🐣 (@swyx) 6 avril 2024
It’s clear that we are at the Inflection point in:
– TTS
– STT
– private LLMs
– (low latency versions of all the above)
– >24h battery life BLE microphones
– Long context/structured… https://t.co/gbRgTTF4YZ pic.twitter.com/pBuKo7iTsMJust hosted the world’s first Personal AI meetup! (where recording is opt out instead of opt in) It’s clear that we are at the Inflection point in:
– TTS
– STT
– private LLMs
– (low latency versions of all the above)
– >24h battery life BLE microphones
– Long context/structured -
OpenAI’s New Role: API and Fine-tuning Engineering Leadership
By
–
(we knew that this was already a title at OAI, and this is more training-heavy, but still the core skills of driving value w/ API, and finetuning/evaling makes it the most in-demand engineering job of the decade) excited for the OpenAI keynote this June
-
Haiku Model Becomes Default Choice in Application Calls
By
–
Nearly every call with this uses Haiku by default 🙂
-
Claude Researcher: Constrained Agent for Detailed Report Generation
By
–
`claude-researcher` is a constrained agent — meaning its behavior is highly-controlled, leading to better results than open-ended agents. It chains together lots of Claude 3 calls (and Google access) that work together to create a detailed report on a topic of your choice.
-
Anthropic Team Releases AI Demo with Alex Albert
By
–
The Anthropic team came through. Demo incoming. Shoutout to @alexalbert__ 🙂
-
John Carpenter’s Dark Star: 50 Years Later and AI Dialog
By
–
John Carpenter's first movie ($60k budget) was released 50 years (and a few days) ago. Dark Star. "Talk to the bomb." "Teach it phenomenology." Echoes of 2001. The bomb's dialog could today be generated by an LLM, I think, coupling together sort of related ideas from philosophy
-
Best Open Source LLM for Fine-tuning: Results from 700+ Models
By
–
Which #opensource LLM provides best results when fine-tuned? We #finetuned 700+ #LLMs so you don't have to! Join our webinar to learn: Which OSS model reigns supreme? How to get GPT4 performance w/ OSS LLMs at a fraction of the cost and more! https://
pbase.ai/3PLRTGT