Neat idea: Database of open STEM problems. Maybe we can add a "Solve Now" button that lets people send the problem to their favorite AI. Then we can track where AIs help make progress. Maybe little progress now, but probably increasing over time.

Working in AI has shown me that most people struggle to give clear, unambiguous instructions with good background, objectives, and success criteria. It's often not the AI but its direction that causes failure.

Yup, here we go, Sora credits to cover generation costs and also establish a new entertainment economy: "we imagine a world where rightsholders have the option to charge extra for cameos of beloved characters and people. we will soon pilot monetization..." Prediction: This will,

Phenomenal case here and further evidence that GPT-5 Pro, working with experts, can make meaningful contributions to science. In this case, it independently identified that dupilumab could prevent a serious food reaction, something that was never previously published, but was

Good and bad news for Canadian AI: Frontier R&D capabilities (about 9 months behind the US) and high AI diffusion. However, we're quite dependent on foreign compute capacity. We really need to leverage our strengths in land and energy to improve here.

When ChatGPT Pulse nails something, it’s magical. I’m about to finish a book. It knows that, and it knows what I’ve read before and what I’m interested in. Today, I got two interesting book recommendations. I didn’t ask, they just showed up. If there was a buy button, I’d

Umm... these are all IMO-level questions or harder, and the best model gets >50%. And this is being positioned as LLMs still struggling in high school math? This isn't, like, grade 10 algebra. These are IMO-level questions.

Also, this is a chart crime. The axis starts at 20, which makes the models look way worse on this new benchmark. They should be past the half way mark on the chart. Seems like everything about this paper was spun to make the models look bad. Meantime, it's >50% on novel IMO-level

Okay, Sora characters are going to be a big deal. The consistency of the character visuals AND voices in this video is great. I didn't notice any jarring discontinuities. I'm sure this still took a lot of work, but way, way, way less work than it would have taken a year ago.

.Nathaniel Whittemore really think on the latest AI Daily Brief episode you should have pointed out that RLI may underestimate AI’s automation potential because tasks AI can do are no longer represented on job boards. For example, research showed a reduction of 21% in writing and coding

So, models don't appear to be deceiving us when they talk about self-awareness. It's the opposite. They're deceiving when they say they don't have it. That's the implication of this study. They turn deception up or down and find that LESS deception = MORE self-awareness claims.

More interesting discussion of the deception-linked self-awareness paper. Training models to say they’re not self-aware may in fact train them to be deceptive. So, we end up with this choice: 1. Acknowledge that the models are self-aware in some way, or “think” they are, and let