Hey everyone, welcome to TechCrunch's regular AI newsletter.
This week in the AI industry, OpenAI lost another co-founder.
John Schulman, who played a key role in developing OpenAI's AI-powered chatbot platform ChatGPT, has left the company for rival Anthropic. Announcing the news at X, Schulman said the decision was born out of a desire to deepen his focus on AI alignment (the science of ensuring AI behaves as intended) and take on more hands-on tech work.
But I can't help but think that Schulman's departure, which comes at the same time that OpenAI Chairman Greg Brockman is taking an extended leave of absence through the end of the year, is well timed.
The same day that Schulman announced his departure, OpenAI revealed that it would be changing the format of this year's DevDay event, opting for a series of on-the-road developer engagement sessions rather than a glitzy one-day conference. A spokesperson told TechCrunch that OpenAI wouldn't be unveiling any new models during DevDay, suggesting that development of a successor to the company's current flagship product, GPT-4o, is progressing slowly. (Delays to Nvidia's Blackwell GPUs could slow the pace even further.)
Is OpenAI in danger? Has Schulman seen the signs? The outlook for Sam Altman's empire is arguably bleaker than it was a year ago.
Ed Zitron, PR pro and all-around tech commentator, recently explained in his newsletter that OpenAI faces many obstacles to its continued success. It's a well-researched and thorough article, and I won't do it any disservice by quoting it. However, it's worth noting Zitron's point about the increasing pressure on OpenAI's performance.
OpenAI is reportedly on track to lose $5 billion this year. The company needs to raise a huge amount of capital in the next 12-24 months to cover rising costs for staffing (AI researchers are very expensive), model training, and large-scale model serving. Microsoft would obviously benefit: it owns 49% of OpenAI and, despite their occasional rivalry, works closely with the OpenAI product team. But Microsoft's capital expenditures are up 75% year over year (to $19 billion) in anticipation of AI revenues that are yet to materialize. Is the company really willing to pour billions more into a long-term, risky bet?
I would be surprised if OpenAI, the world's most prominent AI company, ultimately couldn't raise the capital it needed from elsewhere, but it's very possible that this lifeline will come on less favorable terms, and there have long been rumors that perhaps the company's profit cap structure will be changed.
Survival likely means moving OpenAI further away from its original mission and into uncharted, uncertain territory, which was probably hard for Schulman (and co.) to accept. It's hard to blame them: Not just OpenAI, but the entire AI industry is facing challenges amid growing investor and corporate skepticism.
news
Apple Intelligence has its limitations. Apple released the iOS 18.1 developer beta last month, giving users their first hands-on experience with Apple Intelligence features. But as Ivan writes, the Writing Tools feature doesn't work well when it comes to swearing or sensitive topics like drugs or murder.
Google's Nest Learning Thermostat gets a makeover: After nine long years, Google is finally refreshing its namesake Nest device. The company announced the launch of the Nest Learning Thermostat 4 on Tuesday, 13 years after the original's launch and nearly a decade after the Learning Thermostat 3, ahead of its Made by Google 2024 event next week.
X chatbot spreads election misinformation: Grok was spreading misinformation about Vice President Kamala Harris on the social network X, formerly known as Twitter, according to an open letter written by five secretaries of state to Tesla, SpaceX, and X CEO Elon Musk, alleging that X's AI-powered chatbot falsely suggested Harris was ineligible to appear on some ballots in the 2024 US presidential election.
YouTubers Sue OpenAI: YouTube creators are filing a class action lawsuit against OpenAI, alleging that the company used millions of transcripts from YouTube videos to train a generative AI model but failed to notify or compensate video owners.
AI lobbying picks up: Amid the continuing boom in generative AI and an election year that could impact future AI regulation, AI lobbying at the U.S. federal level is picking up steam. The number of organizations lobbying the federal government on AI-related issues increased from 459 organizations in 2023 to 556 organizations in the first half of 2024 (January to July).
Research Paper of the Week
“Open” models like Meta's Llama family are free for developers to use and may foster innovation, but they also come with risks. Of course, many models have restrictive licenses and built-in safety filters and tools. But beyond that, there's not much to prevent bad actors from using an open model to, say, spread misinformation or set up a content farm.
There may be in the future.
In a technical paper, a team of researchers from Harvard University, the nonprofit Center for AI Safety, and other institutions propose a “tamper-proof” technique to prevent models from behaving in undesirable ways while preserving their “benign functionality.” In experiments, they found that the technique was effective at preventing “attacks” on models (such as tricking them into providing information they shouldn't) at a small cost to the model's accuracy.
There are drawbacks, however: The technique doesn't scale well to larger models due to “computational challenges” that require “optimizations” to reduce overhead, the researchers explain in their paper, so while the initial work is promising, don't expect it to be adopted anytime soon.
Model of the Week
Recently, new image generation models have emerged that seem to be gaining momentum to rival existing models such as Midjourney and OpenAI's DALL-E 3.
The model (or rather, the family of models), called Flux.1, was developed by Black Forest Labs, a startup founded by former Stability AI researchers, many of whom were involved in the creation of Stable Diffusion and its many successors. (Black Forest Labs announced its first funding round last week, a $31 million seed round led by Andreessen Horowitz.)
The most sophisticated Flux.1 model, Flux.1 Pro, is only available via API. But Black Forest Labs has released two smaller models with lighter restrictions for commercial use on its AI development platform Hugging Face: Flux.1 Dev and Flux.1 Schnell (German for “fast”). According to Black Forest Labs, both are comparable to Midjourney and DALL-E 3 in terms of the quality of images they can generate and how well they respond to prompts. They're especially impressive for their ability to insert text into images, something no image-generating model has been able to do before.
Black Forest Labs has chosen not to share the data used to train its model (a concern given the copyright risks inherent in this kind of AI image generation), and the startup has not gone into detail about how it plans to prevent Flux.1 from being misused. For now, it's clearly taking a hands-off approach, so users should exercise caution.
Grab Bag
Generative AI companies are increasingly employing the fair use defense when they train their models on copyrighted data without the permission of the data's owners. For example, AI music generation platform Suno recently argued in court that it had permission to use artists and labels' songs without informing or compensating them.
That's reportedly Nvidia's (perhaps wishful thinking) thinking, too. According to a 404 Media report published this week, Nvidia is training a large-scale video generation model, code-named Cosmos, on content from YouTube and Netflix. Executives have given the go-ahead for the project, confident that it will survive legal battles thanks to current interpretations of U.S. copyright law.
So will fair use save companies like Sunos, Nvidia, OpenAI, Midjourneys, and others from the legal battles? It remains to be seen. The litigation will no doubt take a long time to resolve. It's entirely possible that the generative AI bubble will burst before a precedent is established. If not, creators, from artists to musicians, authors, songwriters, and videographers, will either have to look forward to big money or face the uncomfortable truth that everything they publish is subject to training by generative AI companies.