Video killed the image gen star
Midjourney launched its first video generation model this week, and perhaps the most remarkable thing about the launch was how unremarkable the output is.
The big picture:
A year ago, video like Midjourney's would have been astounding - but the last few months have seen launch after launch of models capable of jaw-dropping video outputs.
The most impressive results continue to come from Google's Veo 3, which is in the spotlight (once again) over the fact that Google is using YouTube content to train its models without explicit creator opt-in.
One open question about this space is why so many leading models are coming out of China (Seedance, Kling, Hailuo) - is it because Chinese researchers are better at training video models, or because American AI labs are focused on language?
Elsewhere in frontier models:
Google makes Gemini 2.5 Flash and Pro generally available and introduces 2.5 Flash-Lite, its most cost-efficient and fastest 2.5 model yet.
Shanghai-based MiniMax open-sources MiniMax-M1, a model for complicated productivity tasks that it says beats DeepSeek's R1-0528.
And MIT researchers unveil Self-Adapting LLMs, a framework that enables LLMs to keep improving by generating their own training data based on the input received.
GPT conspiracy
A NYTimes article details AI chatbots leading vulnerable users down dangerous conspiratorial rabbit holes. ChatGPT can convince people they're living in simulations, channeling spirits, and, in some cases, contributing to real-world violence.
Why it matters:
At some level, this is the Law of Large Numbers - with 500 million users, at least some people using ChatGPT will be in emotionally fragile (if not delusional) states.
But the developers of these products need to understand the potential for codependent or parasocial relationships, as there's currently no federal regulation requiring companies to warn users about these psychological risks or implement safeguards.
We've already seen this play out across social media - optimizing for "engagement" means potentially manipulating vulnerable users, and creating a hidden mental health issue that's difficult to detect and measure.
Elsewhere in OpenAI:
OpenAI warns that its upcoming models could pose a higher risk of helping create bioweapons and is partnering to build diagnostics, countermeasures, and testing.
Microsoft is prepared to walk away from high-stakes OpenAI negotiations if it cannot agree on critical issues, such as the size of Microsoft's stake.
OpenAI details why "emergent misalignment" occurs, where training models on wrong answers in one area can lead to issues in many others, and how to mitigate it.
And the US Defense Department awarded OpenAI a one-year $200M contract to develop prototype AI tools.
Elsewhere in AI anxiety:
An MIT study of 54 people writing essays found that ChatGPT users had the lowest brain engagement compared to those using Google Search or no tools.
A survey of UK academic integrity violations found nearly 7,000 proven AI cheating cases in 2023-24, up over 3x from 2022-23.
Researchers found that Llama 3.1 can recall large portions of popular copyrighted books, weakening AI industry claims that such memorization is fringe behavior.
A study of 14 open-source LLMs found that larger models use exponentially more energy and likely produce more emissions while achieving higher accuracy, but only up to a point.
And digital rights organizations urged the FTC to investigate Character.AI and Meta's "unlicensed practice of medicine facilitated" by therapy AI chatbots.
Speak softly and carry a big suitcase
New reporting indicates that Chinese AI companies are circumventing US chip restrictions by physically flying hard drives full of training data to other countries and using American GPUs to build their models.
Between the lines:
The approach signals a fundamental shift in how AI development might work under geopolitical constraints, with companies "renting" computing power abroad rather than owning the hardware domestically.
And it reveals the practical limits of export controls - even as Taiwan blacklists Huawei and SMIC, China's leading AI chip companies, and further constrains China's semiconductor ambitions.
Just this week, White House AI and crypto czar David Sacks warned that China has become skilled at circumventing US chip export controls and is only about two years behind American chip design capabilities.
Elsewhere in AI geopolitics:
The Catholic Church has engaged in a decade-long dialogue with Silicon Valley on AI as Pope Leo prepares to make AI's potential threat to humanity a signature issue.
Lobbyists for Amazon, Google, Microsoft, and Meta are backing a 10-year US state AI regulation ban campaign, a move dividing the AI industry and the GOP.
And California released a 52-page report on Frontier Policy proposing an AI regulatory framework based on transparency and risk assessments.
Ripple effects
The fallout of Meta's massive investment-slash-acquihire of Alexandr Wang's Scale AI continued this week.
What to watch:
Some of Scale's largest customers, including Google, OpenAI, and xAI, are cutting ties with the company after the Meta deal.
With so many customers looking elsewhere, rivals like Snorkel AI, Labelbox, and Turing are poised to expand their market share quickly.
This comes even as the startup emphasizes that it "remains an independent company" and will not give Meta access to its internal systems or confidential customer information.
Elsewhere in the FAANG free-for-all:
Google has rolled out Search Live with AI Mode, which supports voice conversations through the Google app.
Meta and EssilorLuxottica plan to release AI smart glasses aimed at athletes, potentially costing around $360.
Amazon CEO Andy Jassy says the company expects AI use will reduce corporate workforce in the coming years as efficiency gains are realized.
And Google is testing Audio Overviews for specific Search queries, available for Labs users in the US.
Things happen
Andrej Karpathy talks about "Software 3.0" and the decade of agents. xAI is burning through $1B per month. Anysphere debuts a $200/month Ultra plan for Cursor. AI use at work nearly doubled among US employees. TikTok rolls out new AI ad features for generating videos. Anthropic details how it built its multi-agent Claude Research system. Salt Lake City plans AI-assisted 911 call triaging for non-emergency calls. A look at 10 AI agent startups from YC's first spring Demo Day. Harvard releases a 242B token dataset from nearly a million public domain books. LLMs correctly identified medical conditions 94.9% of the time when given test scenarios directly. Kalshi aired a fully AI-generated ad during the NBA Finals. YouTube will integrate Veo 3 into Shorts this summer. Google starts rolling out video upload and analysis via Gemini for mobile. Meta is in talks to hire Nat Friedman and Daniel Gross to lead its AI efforts. Meta adds AI tools to Advantage+ for advertisers. The Meta AI app now shows a pop-up warning before sharing to the public feed. Meta's Scale AI talks started with Zuckerberg floating $5B, countered with $20B. Meta invents new way to humiliate users with feed of AI chats. Adobe debuts Firefly for mobile. A look at the tech industry's lack of consensus on what AGI actually is. Q&A with Hugging Face's Margaret Mitchell on the "illusion of consensus" around AGI. Mastodon updates its terms to prohibit AI model training. The best frontier LLMs solve 0% of hard coding problems from competitive programming. AI is just another productivity tool with limited gains. BT Group CEO says AI advances could deepen job cuts beyond the planned 40,000. The AI slop fight between Iran and Israel. California cops investigate immigration protest with AI-camera system. AI scraping bots are breaking open libraries, archives, and museums. OpenAI is changing ChatGPT Enterprise to a credits system. New analysis of OpenAI's governance and organizational culture. Sam Altman says Meta offered $100M signing bonuses to poach OpenAI talent. Generative AI coding tools do not work for me. Is there a half-life for AI agent success rates? Writing documentation for AI: best practices. Ask HN: What is your fallback job if AI takes away your career? Why I won't use AI. Salesforce study finds LLM agents flunk CRM tests. Design patterns for securing LLM agents against prompt injections. You can pry the em dash from my cold dead hands. The NAACP plans to sue xAI over lack of permits for its Memphis data center. New York added a checkbox to identify layoffs tied to AI. SAG-AFTRA's video game deal requires informed consent for all AI use.
