Join us at the @EmbVisionSummit in Santa Clara, USA, from May 21 – 23! Witness our cutting-edge AI solutions firsthand. Bram Verhoef, our head of Machine Learning, will present on May 22 about "Axelera AI: Digital In-Memory-Compute for Fast, Energy-Efficient Computer Vision"
RESEARCH
-
Evaluating Speed Cost and Offline AI Performance Metrics
By
–
Speed, cost and the ability to run offline on a personal device are the things I care most about I still don't have a great answer for how to best evaluate these things but I'm researching that at the moment
-
Google’s Machine Learning Energy Costs: Fact or Fiction?
By
–
> Google alone spends 75% of its electrical cost only on machine learning today i've read this somewhere, but I'm wondering if there's a reference source for this claim?
-
Historical Technology Errors and Intellectual Property Practices Examined
By
–
Apparently new information technologies have been beset by egregious errors and dubious IP practices for at least half a millennium.
-
GPT-5, Gemini 2.0, and the race for LLM advancement
By
–
I see a lot of comments on how much is riding on GPT-5 to show LLMs haven’t plateaued (I have said similar). But it isn’t like OpenAI is the only company expecting future big model improvement. We know Gemini 2.0 is in the cards, and we know Anthropic & X are training new models.
-
Real-world AI attacks remain limited beyond screenshots and prompt leaks
By
–
There have been a bunch of research/demo attacks, but I haven't seen a real-world malicious attack that caused actual damage yet – beyond producing embarrassing screenshots or leaking system prompts
-
OpenAI releases detailed prompt injection evaluation paper
By
–
New paper from @OpenAI on prompt injection – it's the most detailed evaluation of the problem I've seen from them so far, and has some very interesting details Posted some of my notes on the paper on my log here: https://
simonwillison.net/2024/Apr/23/th
e-instruction-hierarchy/
… -
Experimenting with thinking in other languages for AI research
By
–
Thinking in other languages is a fascinating idea. I want to experiment with that. Very clever.
-
Training costs for 8B parameter models relatively affordable
By
–
Couple hundred. Pretty cheap b/c only 16K on 8B
-
70B Model Context Window Limitations and Hardware Requirements
By
–
Ooof. I don't know if 128K is possible on the 70B, unless someone can get me a rig with like 16 or more A/H100s. 32K is likely doable though.