AI News: Self-Learning Models and AI in Advertising

TL;DR

Google DeepMind's AlphaEvolve is a self-improving AI that evolves code using Gemini Flash for idea breadth and Gemini Pro for critical depth, and it discovered a faster algorithm for multiplying 4x4 complex matrices, beating a record set in 1969. Meanwhile, Meta, Netflix, and YouTube are embedding AI across advertising to auto-generate and optimally place ads.

Transcript

It's been another really interesting week in the world of AI. We got new models that train themselves. AI infiltrating every part of advertising, including what AI will look like in the apocalypse, plus some new chat GPT features and probably the best dance moves you'll ever see coming from a humanoid robot. So, don't want to waste your time. Let's... Read More

Key Insights

• AlphaEvolve is a self-improving AI from Google DeepMind that evolves code rather than just writing it, using Gemini Flash to maximize the breadth of ideas explored and Gemini Pro to provide critical depth and insightful evaluation of which ideas are viable solutions.
• AlphaEvolve verifies, runs, and scores proposed programs using automated evaluation metrics that give an objective, quantifiable assessment of each solution's accuracy and quality, and Google DeepMind uses it internally to optimize data centers, TPU circuit design, and Gemini model training.
• AlphaEvolve discovered a new algorithm to multiply 4x4 complex-valued matrices, improving on the previous best-known method created in 1969, meaning it beat a math record unbeaten for 56 years by inventing new problem-solving methods rather than reusing known ones.
• Absolute Zero is a new RLVR training paradigm from Tsinghua University, Beijing Institute for General Artificial Intelligence, and Penn State where a single model proposes its own tasks to maximize learning and improves reasoning by solving them without any external data.
• The Absolute Zero Reasoner self-evolves its training curriculum by using a code executor to validate proposed code reasoning tasks and verify answers, achieving state-of-the-art coding and math performance despite using no human-curated examples, though it only works for math and code.
• Meta's future advertising vision, described by Mark Zuckerberg, lets any business state an objective, say what they will pay, and connect their bank account, after which Meta delivers as many results as possible, potentially creating the ad creative and targeting automatically.
• Netflix is using generative AI in a new modular ad framework that blends spots into its shows, placing product images over backgrounds inspired by IP like Stranger Things, Wednesday, and Bridgerton, and can insert these ads mid-content or when a viewer hits pause.
• YouTube announced Peak Points at its Brandcast event, a Gemini-built product that identifies the most meaningful or peak moments within popular content to place brand ads where audiences are most engaged and least likely to click away.

Questions & Answers

Q: What is AlphaEvolve and how does it work?

AlphaEvolve is a self-improving AI from Google DeepMind that evolves code rather than just writing it. It uses two Google models: Gemini Flash maximizes the breadth of ideas explored by brainstorming many potential solutions, while Gemini Pro provides critical depth with insightful suggestions to evaluate which ideas are viable. AlphaEvolve then verifies, runs, and scores the proposed programs using automated evaluation metrics that provide an objective, quantifiable assessment of each solution's accuracy and quality.

Q: Why is AlphaEvolve's matrix multiplication discovery significant?

AlphaEvolve discovered multiple new algorithms for matrix multiplication, including one to multiply 4x4 complex-valued matrices. The previous best-known algorithm for that specific problem was created in 1969, so AlphaEvolve found a better way where none had been found in 56 years. This matters because it invents genuinely new problem-solving methods rather than reusing methods already known and scraped from the internet, which the video frames as a major breakthrough.

Q: What does Google DeepMind use AlphaEvolve for internally?

Google DeepMind has been using AlphaEvolve internally to help optimize several of its own systems. According to the video, these include optimizing their data center, their TPU circuit design, and even helping to train the Gemini models themselves. Beyond internal optimization, AlphaEvolve can also propose new approaches to complex mathematical problems, such as discovering the improved matrix multiplication algorithms mentioned in the announcement, which came the week before Google IO.

Q: What is the Absolute Zero AI training method?

Absolute Zero is a new RLVR paradigm from a research paper by Tsinghua University, the Beijing Institute for General Artificial Intelligence, and Penn State. A single model learns to propose tasks that maximize its own learning progress and improves reasoning by solving them without relying on any external data. The Absolute Zero Reasoner self-evolves its training curriculum using a code executor to validate proposed code reasoning tasks and verify answers, serving as a unified source of verifiable reward.

Q: Does Absolute Zero lead to AGI?

No. The video stresses that Absolute Zero only excels at math and code, so claims that it is a path to AGI are not really accurate. For something to be AGI, it would need world knowledge and the ability to solve any type of problem. A model like this cannot give business advice like ChatGPT, nor answer questions about world history or current events. It specializes in creating math and coding problems, solving them, and verifying results with a code executor, and does not work outside that domain.

Q: How does Mark Zuckerberg envision the future of Meta advertising?

Mark Zuckerberg, interviewed by the president of Stripe, described advertising where any business can state its objective, such as getting new customers or selling products, say how much they are willing to pay for those results, and connect their bank account. Meta then delivers as many results as possible. In the future, a small or large business may not even need to start with a creative; it could just bring a goal and a budget, and Meta would deliver results across its platform using AI to optimize.

Q: How is Netflix using AI in its advertising?

Netflix is using generative AI to make ad breaks look less like ad breaks. According to The Verge, its new modular ad format lets advertisers make spots that blend in with Netflix shows and movies. At its Upfront event, Netflix demonstrated placing a product image over a background inspired by a show like Stranger Things. On ad-supported plans, viewers might see ads blending into shows like Bridgerton or Wednesday, inserted mid-content or shown when the viewer hits pause.

Q: What is YouTube's Peak Points ad feature?

Peak Points is a new advertising product YouTube announced at its Brandcast event, built with Gemini. It identifies the most meaningful or peak moments within YouTube's popular content to place a brand where audiences are the most engaged. Essentially, when a video gets really good and viewers are leaning in and invested, the AI recognizes they are unlikely to click away, and places the ad at that exact moment of peak engagement to maximize the brand's exposure.

Summary & Key Takeaways

• AlphaEvolve from Google DeepMind is a self-improving AI that evolves code using Gemini Flash to brainstorm many ideas and Gemini Pro to critically evaluate them, then verifies and scores solutions with automated metrics. It powers internal work like data center and TPU optimization.

• AlphaEvolve discovered new matrix multiplication algorithms, including one for 4x4 complex-valued matrices that beat a 1969 record. Separately, the Absolute Zero paper introduced a model that generates its own math and coding problems, solves them, and verifies answers with a code executor, needing no external data.

• AI is infiltrating advertising: Meta lets businesses state a goal and budget while it delivers results, Netflix uses generative AI to blend ads into show worlds like Stranger Things and Wednesday, and YouTube's Peak Points uses Gemini to place ads at the most engaging moments.