AI Reading for Saturday February 22

Feb 22, 2025

GPT-5 maybe May around the time of Microsoft's Build conference, which lines up with Sam Altman's 'months' comment (vs GPT-4.5 in 'weeks') - Firstpost

OpenAI presentation shows big increase in revenue, shift in compute spend to Stargate and Softbank - The Information

OpenAI bans accounts working on disinformation, surveillance campaigns. - Axios

This makes it sounds like someone was making a Llama-based tool to identify China-unfriendly social media posts, used OpenAI to debug the code, and OpenAI banned them because of code they were writing to use Llama, as opposed to abusing OpenAI. - NY Times

If I'm using OpenAI to write malicious security exploits, refusal seems appropriate. Or if I'm writing code to use Llama for automated hacking. Analyzing posts for sentiment about China and the CCP, is that a crime? Is it being used to physically target people in real time, to harass, spam or censor them? fine line

New Alexa announcement expected next week (if no more delays) - CNET

Amazon values its $8b Anthropic investment at $14b - Business Insider

Is AI really thinking and reasoning — or just pretending to? - Vox

Deep research may just mean more AI slop (derp research?) - Intelligencer

Genspark uses many AI models to do do deep research, compete with offerings from Google, OpenAI, others - Reuters

The Humane Ai Pin Will Become E-Waste Next Week. Sorry early adopters. - WIRED

Alibaba in quest for AGI.

no word on the 40 thieves

Why AI Spending Isn’t Slowing Down - WSJ

Apple announce VisionOS update with AI - Apple Newsroom

Google, Meta execs blast Europe over strict AI regulation - CNBC

Figure demos vision-language-action model allowing you to tell the robot to do something - FigureAI

Cool demo, but these things seem 10 years away from product/market fit, other than as novelties, in industrial settings. on a longer timeline ofc humans are screwed.

Tough but fair: Why would anyone want a humanoid robot?

I think better robot vacuums, static appliances, ability to talk to realistic human avatars on a screen for 10 years at least before widespread humanoid robots.

There is something about talking to a real robot though… the human mind is programmed to pay attention to human figures. like the flailing tube men at car washes. Or how we see human faces in cars, and are super alert to pairs of eyes looking at us, from evolving as prey for stealthy sabretooth tigers.

Some books on Big Tech's impact on political economy and geopolitics. Thinking people are totally lost while barbarians, scoundrels, and plunderers seize the commading heights, the means of production of shared narratives, manufactured consent, and social control

"Civilization of the Mind in cyberspace" my ass.

Writing, and teaching writing, in an age where AI can write - WSJ

Richard Dawkins asks ChatGPT if it is conscious - Richard Dawkins

AI coding might reduce code re-use, increase technical debt. Why look up a function you already wrote, or make sure it supports all use cases, when it's so easy to ask the AI to write a code to do exactly what you want? - GitClear

100x improvements are sometimes just bad evals - TechCrunch

Good AI Engineer Summit last week in NYC, will fast-forward through some great talks I missed www.ai.engineer/summit/2025 - AI.engineer

Strategy track - YouTube

Dev track - YouTube

10 Mistakes Beginning AI Practitioners Make That You Should Avoid. - Hacker Noon

First thing is, start with a baseline naive model to get a feel for how much a complex model might add ... better data, better features and better understanding of the business generally beats better math and deep learning.

Llama 3 from scratch - GitHub

Follow the latest AI headlines via SkynetAndChill.com on Bluesky

AI Reading for Saturday February 22

Discussion about this post