Hello everyone, welcome to TechCrunch's regular AI newsletter. If you'd like to have this sent to your inbox every Wednesday, sign up here.
On Monday, Anthropic CEO Dario Amodei sat down with AI influencer Lex Fridman for a five-hour podcast interview. The two covered a wide range of topics, from the timeline of superintelligence to the advancement of Anthropic's next flagship technology.
To save you the trouble of downloading, we have extracted the important points.
Despite evidence to the contrary, Amodei believes that “scaling up” models is a viable path to more capable AI. Amodei clarified that scaling up means increasing not only the amount of compute used to train the model, but also the size of the model and the size of the model's training set.
“Scaling is probably going to continue, and there's a magic to it that we haven't yet been able to explain theoretically,” Amodei said.
Amodei also doesn't think a lack of data poses a challenge to AI development, unlike some experts. By generating synthetic data or extrapolating from existing data, AI developers can “work around” data limitations, he says. (I'm still not sure if the synthetic data problem is solvable, but I'll note it here.)
Amodei acknowledged that the cost of AI computing could become even higher in the short term, due in part to scaling effects. He predicts that companies will spend billions of dollars on clusters to train models next year, and hundreds of billions by 2027. (In fact, OpenAI is rumored to be planning a $100 billion data center.)
And Amodei said candidly that even the best models are inherently unpredictable.
“It's very difficult to control the behavior of a model, to control the behavior of a model in all situations at once,” he said. “There's a 'whack-a-mole' aspect where if you keep pushing on one thing, other things start moving as well that you don't notice or can't even measure.”
Still, Amodei predicts that Anthropic, or a rival, will create a “superintelligent” AI that can surpass “human-level” performance on many tasks by 2026 or 2027. And he is concerned about the implications this will have.
“We're rapidly running out of really compelling disincentives, really compelling reasons why this won't happen in the next few years,” he said. “I'm worried about the economy and the concentration of power. Actually, what I'm more worried about is the abuse of power.”
So it's good that he's in a position to do something about it.
news
An AI news app launched by a former Twitter engineer: AI newsreader Particle aims to help readers better understand the news with the help of AI technology.
Writer raises funding: Writer raises $200 million at a $1.9 billion valuation to scale its generative AI platform for enterprises.
Build on Trainium: Amazon Web Services (AWS) has launched Build on Trainium, a new program that will award $110 million to institutions, scientists, and students researching AI using AWS infrastructure.
Red Hat acquires startup: IBM's Red Hat is acquiring Neural Magic, a startup that optimizes AI models to run faster on general-purpose processors and GPUs.
Free Grok: X (formerly Twitter) is testing a free version of the AI chatbot Grok.
Grammys powered by AI: The Beatles song “Now and then,” released last year, was nominated for two Grammys after being refined using AI.
Anthropic for Defense: Anthropic is partnering with data analytics company Palantir and AWS to provide U.S. intelligence and defense agencies with access to Anthropic's Claude family of AI models.
New Domain: OpenAI acquires Chat.com and adds it to its collection of high-profile domain names.
This week's research paper
Google claims to have developed an improved AI model for flood prediction.
The model builds on the company's previous research in this area and can accurately predict flood conditions in dozens of countries up to seven days in advance. In theory, the model could provide flood predictions anywhere on the planet, but Google notes that many regions lack historical data to test it.
Google offers a waiting list for API access to our models for disaster management and hydrology professionals. We also make predictions from models available through the Flood Hub platform.
“We want to contribute to the research community by making our predictions available globally on Flood Hub,” the company said in a blog post. “Professional users and researchers can use these data to conduct more research and analysis on how flooding impacts communities around the world.”
this week's model
AI developer Rami Seid has released a Minecraft simulation model that can run on a single Nvidia RTX 4090.
Similar to the “open world” model recently released by AI startup Decart, Seid's model, called Lucid v1, emulates the Minecraft game world in real time (or close to it). Lucid v1 has 1 billion parameters, captures keyboard and mouse movements, generates frames, and simulates all physics and graphics.
Output from Lucid v1 model. Image credit: Rami Said
Lucid v1 has the same limitations as other game simulation models. The resolution is very low and you tend to “forget” the level layout quickly. Rotate your character to see the rearranged scene.
But Seid and her partner Ollin Boer Bohan say they plan to continue developing the model, which can be downloaded and enhanced with an online demo here .
grab bag
DeepMind, Google's leading AI lab, has released the code for AlphaFold 3, an AI-powered protein prediction model.
AlphaFold 3 was announced six months ago, but DeepMind caused controversy by withholding the code. Instead, it provided access through a web server that limited the number and type of predictions scientists could make.
Image credit: Google DeepMind
Critics saw the move as an effort to protect DeepMind's commercial interests at the expense of reproducibility. DeepMind spinoff Isomorphic Labs is applying AlphaFold 3, which can model proteins in conjunction with other molecules, to drug discovery.
Scientists can now use this model to make any predictions they like, including how proteins will behave in the presence of potential drugs. Scientists affiliated with academic institutions can request code access here.