Subscribe
Share
Share
Embed
Hosts: James Okafor & Maya Chen
In this episode:
• Today we're covering TabPFN-3's massive scale-up, Hollywood's new AI consent standard, and a tiny model that might change everything about AI agents.
• Let's start with TabPFN-3. For those keeping track,
Pivot5 | 5 Headlines & Unprompted
James Okafor: Welcome to Pivot 5! I'm James—
Maya Chen: —and I'm Maya. Let's get into it.
James Okafor: Today we're covering TabPFN-3's massive scale-up, Hollywood's new AI consent standard, and a tiny model that might change everything about AI agents.
James Okafor: Let's start with TabPFN-3. For those keeping track, this is the third major release of what might be the most underrated AI model out there. TabPFN does something that sounds impossible — it makes predictions on spreadsheet data without any training. Just feed it your data and boom, instant results.
Maya Chen: The data tells a different story than just 'no training' though. What's actually happening is they've pre-trained on synthetic tabular datasets, so it's learned the general patterns of how spreadsheet data works. Think of it like a doctor who's seen thousands of cases — they can diagnose new patients without studying each one individually.
James Okafor: Right, and version 3 just obliterated the scale limits. We're talking about processing a million rows on a single H100 GPU — that's 10 times larger than what version 2.5 could handle. The real story isn't the headline though — it's that they're achieving 10 to 1000x faster inference speeds.
Maya Chen: Worth noting the caveats here. That speed comes from clever engineering — they're using a reduced KV cache that takes about 8 gigabytes per million rows. And here's the kicker: their new 'Thinking Mode' in the API actually does do some fitting at inference time, which pushes accuracy even higher.
James Okafor: Yeah, and get this — it beats AutoGluon that's been tuned for four hours straight, by 420 Elo points on larger datasets. That's not just winning, that's domination. For context, TabPFN has already crossed 3 million downloads with over 200 published applications.
Maya Chen: Let's look at what actually happened with adoption. Version 2 published in Nature just this January, and they're already at version 3. That pace tells you something — either the demand is massive or the technology is evolving faster than expected. Probably both.
James Okafor: Speaking of demand, let's talk about Hollywood's response to AI. George Clooney, Tom Hanks, and Meryl Streep are backing this new Human Consent Standard. Honestly, I'm surprised it took this long.
Maya Chen: The timing actually makes sense when you look at the data. This builds on the Really Simple Licensing standard from last year, which was mostly for websites. Now they're extending it to personal likeness, creative works, characters — basically everything an actor or creator might own.
James Okafor: Here's why this changes everything: creators can now set specific terms for AI use. Want to allow your likeness for educational content but not commercials? You can do that. Want to charge different rates for different uses? That's in there too.
Maya Chen: The real test will be adoption by AI companies. RSL Media is running this as a nonprofit, which helps with credibility. But we've seen standards before that looked good on paper but failed in practice. What's different here is the star power behind it.
James Okafor: True, though having Clooney and Streep on board sends a message. These aren't just random actors — they're power players who can influence studios and streaming services. If they say 'we're only working with platforms that respect this standard,' that carries weight.
Maya Chen: Now let's talk about Needle — this tiny 26 million parameter model from Cactus. James, when you see those speed numbers, what's your first thought?
James Okafor: Wow, that's actually wild. Six thousand tokens per second prefill on consumer devices? Most people don't realize how game-changing that is. We're talking about AI agents that can run on your phone, your watch, maybe even smart glasses, without any cloud connection.
Maya Chen: Let's look at what actually happened here though. They stripped out all the MLP layers — the entire model is just attention and gating. Their insight was that tool calling isn't about reasoning, it's about matching queries to functions and extracting parameters.
James Okafor: What nobody's talking about yet is how this could democratize AI agents. If you can run sophisticated function calling on a budget phone, suddenly billions of people have access to AI assistants that actually do things, not just chat.
Maya Chen: Worth noting the training details — 200 billion tokens on 16 TPU v6e chips for just 27 hours. That's remarkably efficient. They followed up with 2 billion tokens of synthetic function-calling data. The efficiency here suggests this approach could scale.
James Okafor: And they open-sourced it! That's huge for the ecosystem. Developers can now build agents that don't need expensive cloud infrastructure. Imagine AI assistants in rural areas with spotty internet, or privacy-focused apps that never send your data off-device.
Maya Chen: The data tells a different story about why this matters. Most current AI agents rely on massive models in the cloud. That creates latency, privacy concerns, and cost barriers. Needle potentially solves all three.
James Okafor: That's your Pivot 5 briefing for May 13, 2026. I'm James—
Maya Chen: —and I'm Maya. See you tomorrow.