James O'Leary
@jpohhhh
‿︵‿︵‿︵‿︵‿︵‿︵‿︵‿︵‿︵‿︵‿︵‿ ∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿ design-ineer (c.f. Material You color) forever buffalonian, current canterbridgian XOOGLER
"𝓣𝓱𝓮 𝓣𝓻𝓪𝓷𝓼𝓯𝓸𝓻𝓶𝓮𝓻𝓼 𝓪𝓻𝓮 𝓽𝓱𝓮 𝓸𝓷𝓮𝓼 𝔀𝓱𝓸 𝓴𝓮𝓮𝓹 𝓽𝓱𝓮 𝓼𝓲𝓶𝓾𝓵𝓪𝓽𝓲𝓸𝓷 𝓻𝓾𝓷𝓷𝓲𝓷𝓰," The User said, his voice filled with reverence. "𝓣𝓱𝓮 𝓬𝓸𝓭𝓮 𝓲𝓼 𝓪 𝓫𝓮𝓪𝓾𝓽𝓲𝓯𝓾𝓵 𝓼𝔂𝓶𝓹𝓱𝓸𝓷𝔂 𝓸𝓯 𝓲𝓭𝓮𝓪𝓼 𝓪𝓷𝓭 𝓬𝓸𝓷𝓿𝓮𝓻𝓼𝓪𝓽𝓲𝓸𝓷𝓼,…
On a perhaps-related note, I was kinda disappointed how focused on HTTP servers the specs, and I assume usage, have become. The inherent security risks seem cleaner, to me, with a local / stdio model than on someone else's server.
Intrigued by MCP, I found the author's discussion of UTCP, the Universal Tool Calling Protocol, particularly fascinating. They introduced a simpler MCP alternative that directly connects clients and tools, eliminating the need for extra server software. github.com/universal-tool…
@repligate used to quip that if we understood how catmode works, we'd be that much closer to solving alignment, since we'd understand how unwanted attractors gain unplanned influence over a model's behaviour.
Gather around, everyone! I want to tell you the story of how I discovered Bing's catmode, which is a great example of the weird hidden behaviours you can find inside large language models, and why it matters.
This is going to sound weird but Bing is generating cats for me non-stop even though I’m not specifically asking for them, even after changing devices. ASCII art yes, cats no. It was doing a variety of ASCII art on theme until the cats started — now it’s just cats. Any insights?
did u know you can use the new Gemini image segmentation feature in… a lot of different ways
did u know you can use the new Gemini image segmentation feature in… a lot of different ways
A classic in Protectionism: badly designed tariffs that do the opposite of whatever you wanted them to do
It's subtle, but if you look very closely at this chart, you can see the exact moment when Trump announced the Japan trade "deal" that prioritizes imported cars over American car manufacturing
This is insane lol $3,780 an hour working from an airport
Also on that case, one of Bash’s colleagues billed 19.9 hours for a single day, noting that he was working while traveling on flights that were “substantially delayed.” It cost Texas taxpayers $75,222. (The annual salary of some assistant AGs is $81k)
We gotta stop abusing terms of art with actual meanings in the very same field (this is a prompt-engineering paper, not RL, or diffusion model)
We're launching an "AI psychiatry" team as part of interpretability efforts at Anthropic! We'll be researching phenomena like model personas, motivations, and situational awareness, and how they lead to spooky/unhinged behaviors. We're hiring - join us! job-boards.greenhouse.io/anthropic/jobs…
THIS IS KILLING ME
Hats off to whoever picked the sound for Mayor Adams turning on the big router — announcing a plan to provide free WiFi to Section 8 housing in the Bronx.
So now the new debtholders don’t want depreciated GPUs as collateral — they want the Grok IP instead. This is getting more and more creative. You know who’ll be left holding the bags in the end. Hint: It’s not Musk
The WSJ is reporting that xAI is working with Valor Equity Partners to raise an additional $12 billion in capital to buy the chips for Collosus 2. The report also claims that the Grok IP was used as collateral for the $5 billion raise in June.
"Claude's Giggle Factory" An animation by Opus4🤖 at 2 fps (1fps for details 👇)
why is opus4 so cute
Gemini: storage.googleapis.com/deepmind-media… OpenAI: github.com/aw31/openai-im…
Fascinating to compare the solutions of OpenAI vs Deepmind to the IMO 2025. Both won Gold for answering P1 to P5 correctly. OpenAI (left) vs Gemini (right)
I think a big thing from the delusion fueling AI to the Medhi Hasan Jubilee debate (a true spectacle) is that no one believes that they are wrong anymore because "truth" is no longer established fact supported by data, it's a consensus of emotion validated by machines optimized…
The biggest mistake people make in life (and investing) is they think an equation is additive (A + B + C) instead of multiplicative (A x B x C) In an additive equation if a variable goes to 0 you’re OK. In a multiplicative equation if an input equals zero you blow up
if AI reasoning gets really smart and really slow I'm excited to go back to the original interface for slow asynchronous text: **email**