weekend ai reads for 2025-10-24

Hinton: Now, you just said something that many people say: “This isn't understanding. This is just a statistical trick.”

So let me ask you the question, well, how do you decide what word to say next?

…

So the way you do it is pretty much the same as the way these large language models do it. You have the words you've said so far. Those words are represented by sets of active features. So the word symbols get turned into big patterns of activation of features … and these neurons interact with each other to activate some neurons that go ping, that are representing the meaning of the next word, or possible meanings of the next word. And from those, you pick a word that fits in with those features.

That's how the large language models generate text, and that's how you do it too.

Stewart: And you're saying that can be built into these models as well.

Hinton: That can also be done with pings. That can be done by these neural nets.

Stewart: Whoa. But is the suggestion, then, that with enough data and enough processing power, their brains can function identically to ours? Are they at that point?

Hinton: OK. They're not exactly like us. The point is they’re much more like us than standard computer software’s like us. Standard computer software, someone programmed in a bunch of rules, and if it follows the rules, it does what they expect it to do.

Brains just came from a very different process, and I’m very hesitant to take inspiration from it because we’re not actually running that process. In my post, I said we’re not building animals. We’re building ghosts or spirits or whatever people want to call it, because we’re not doing training by evolution. We’re doing training by imitation of humans and the data that they’ve put on the Internet.

You end up with these ethereal spirit entities because they’re fully digital and they’re mimicking humans. It’s a different kind of intelligence. If you imagine a space of intelligences, we’re starting off at a different point almost. We’re not really building animals. But it’s also possible to make them a bit more animal-like over time, and I think we should be doing that.

In this work, we propose a hybrid model where we activate reasoning mechanisms in base models at the right time to elicit thinking-model-level reasoning chains, implying that thinking models exploit already existing capabilities. … Across three base and four thinking models, using GSM8K and MATH500, our hybrid model recovers up to 91% of the performance gap to thinking models without any weight updates while steering only 12% of tokens.

pretraining on bad data causes lasting “cognitive decline” in LLMs, reducing reasoning, leading to thought-skipping; even fine-tuning can only partially reverse the damage; like we say around here, it all comes back to the data

BDH can be represented as a brain model. The working memory of BDH during inference entirely relies on synaptic plasticity with Hebbian learning using spiking neurons. We confirm empirically that specific, individual synapses strengthen connection whenever BDH hears or reasons about a specific concept while processing language inputs. The neuron interaction network of BDH is a graph of high modularity with heavy-tailed degree distribution. The BDH model is biologically plausible, explaining one possible mechanism which human neurons could use to achieve speech. BDH is designed for interpretability. Activation vectors of BDH are sparse and positive. We demonstrate monosemanticity in BDH on language tasks. Interpretability of state, which goes beyond interpretability of neurons and model parameters, is an inherent feature of the BDH architecture.

Every podcast is better at 2.0 speed!

I see it in my behavior. I’d hazard a guess I visit 10% of the websites I did a year ago.

I recently co-hosted a birthday party with a friend and used Gemini for the invitation. My vision: our younger selves blowing out candles together. I gave Gemini a couple of childhood photos (with my friend’s permission), and it placed us together in front of the same cake. It got our outfits exactly right, and the faces could have fooled our own moms.

There is one area where local models beat any hosted model: End-to-end latency, i.e., the time it takes from your prompt to get the whole response back. …

Time and time again I underestimate the effect of latency and speed when it comes to usability of applications. Having instant reactions to what you do unlocks so many new possibilities as you don't have to sit around awkwardly and wait up to five seconds just for something to appear on your screen.

Personally, I use Claude to draft code for me, including the code that builds this website, and I don’t use AI assistance to write prose. That decision is about what skills I want to keep sharp. I’m fine with not being about to compose code off the top of my head, but I will not outsource my ability to write to an LLM. Skills are like muscles: use ’em or lose ’em.

“Generate 5 responses with their corresponding probabilities, sampled from the full distribution.”

The method, called Verbalized Sampling (VS), helps models like GPT-4, Claude, and Gemini produce more diverse and human-like outputs—without retraining or access to internal parameters.

“Customers are getting their head around how to deploy AI,” Benioff, the CEO of Salesforce, told CNBC host Jim Cramer on Tuesday.

“The speed of innovation is far exceeding the speed of customer adoption,” said Benioff to Cramer. “These customers have to go back and modify massive architectures they have and systems they’re running.”

People worry that not being seen as mindless, uncritical AI cheerleaders will be a career-limiting move in the current environment of enforced conformity within tech, especially as tech leaders are collaborating with the current regime to punish free speech, fire anyone who dissents, and embolden the wealthy tycoons at the top to make ever-more-extreme statements, often at the direct expense of some of their own workers.

At the conference, Griffin said generative AI is unlikely to create widespread changes. It will have impact, but one that won’t be profound, and the technology will disproportionately hit different sectors, he added, according to the people in attendance.

Students at the Pratt School of Design are navigating art and creativity in the age of AI as art schools are seeing a rise in applications. NBC News’ Gadi Schwartz has more on how students are exploring the intersectionality between technology and art.

It’s not really fair that I got to spend a year in my 20s studying a language full-time; that’s an extremely high, expensive barrier to engaging with another society, and it’s nice that other people can do it more easily.

It is extremely important to have a good scaffold. You’ll notice the scores for SWE Bench are much better when there’s a good framework. I’ve open sourced my monorepo scaffold at the end of this thread.

The newest versions of private AI tools like Jan run easily on my 2021 Mac laptop, cost nothing, are easy to use. They’re a good alternative to costlier AI platforms.

Human writers seemed to be losing the battle against AI, with AI-generated articles slightly overtaking human-written articles between November 2024 and March 2025, according to raw data. In January 2025, that data found AI-generated articles peaked at 55%.