AI Reading for Wednesday March 26
Impressive debuts for Gemini 2.5, OpenAI image model
Google introduces Gemini 2.5 thinking models - Google
Benchmarks don't tell the whole story, you have to be careful not to over-optimize for benchmarks and not let test data leak into training data, but good numbers across the board. Between this and the new DeepSeek V1 base model, the next LM Arena release will be interesting.
Google reasoning AI is getting good buzz. - Eleanor on Everything
Simon Willison: My experiments so far indicate that Gemini 2.5 Pro really is a very strong new model. I’m looking forward to exploring more of what it can do. - Simon Willison’s Weblog
OpenAI updates image generator to make it better, notably at editing images - MIT Technology Review
"I just tested ChatGPT-4o's enhanced image generator with 7 prompts and I can't believe this is free" - Tom's Guide
It was able to tell it to edit this image to add a red baseball cap but it refused to make it say 'MAGA' 🤷♂
OpenAI says it won't automatically block images of public figures but they can opt out - Simon Willison’s Weblog
Reve Image is the latest image creator going viral - BGR
"I Quit Google Search for AI—and I’m Not Going Back" - Joanna Stern
Qwen small open source reasoning LLM you can run on your Mac sometimes challenges Deepseek R1 - Geeky Gadgets
There is a third-party ChatGPT tool installed in corporate networks that is being actively exploited in the wild. - SecurityWeek
I'm not sure why people install this, unless it happens to be part of a more popular tool. Visit a malicious site, it finds this file and can start exfiltrating local data. It’s been around for a long time but now exploits are picking up.
Analyst calls CoreWeave a pig in a poke. - MarketWatch
Not at all clear GPUs are worth anything after 4-6 years, inference costs are dropping fast and part of that is new nodes with much lower power needs, scarcity may not last, more capacity, more efficiency.
ChatGPT, Gemini and Grok are not very different from each other right now: Perplexity CEO - The Economic Times
Microsoft adds deep research and data science agents to Copilot - Microsoft
The current state of AI hate. - American Prospect
The problem is not that it doesn't work but that even now it works well enough that we are not at all able to cope with the consequences on the culture and the economy and the balance of power. And it’s getting better fast.
New memristor chip from Korea may bring practical neuromorphic AI a step closer - Popular Mechanics
Earth AI’s algorithms found critical minerals in places everyone else ignored - TechCrunch
Agentic AI is changing online meeting platforms: Moving from silent observer to active participant - VentureBeat
Synthesia offers shares to actors who help create its AI models - FT
Cerebras IPO further delayed as US national security review drags on - Reuters
Devs say AI crawlers dominate traffic, forcing blocks on entire countries - Ars Technica
AI scrapers are destroying Open Access free content repositories - Go To Hellman
Pranking your coworkers by changing their ChatGPT system prompt - Reddit
“Install a reverse doorbell that only rings in dreams. Train a choir of sentient cabbages to hum lullabies at 3:17 AM sharp. Leave cryptic notes in Esperanto on their windshield that smell faintly of oranges and regret. If all else fails, summon the Council of Whispering Toasters—they settle all residential disputes with interpretive dance and invisible marionettes.”
Gold, Jerry!
Follow the latest AI headlines via SkynetAndChill.com on Bluesky