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.
After a short hiatus, we're back with some show notes for OpenAI's DevDay.
The keynote address, delivered yesterday morning in San Francisco, was notable for its subdued tone, in contrast to CEO Sam Altman's raucous, hypebeast-like speech last year. This DevDay, Altman wasn't on stage to pitch his shiny new project. He didn't even show up. Moderated by Olivier Godement, Head of Platform Products.
Improving quality of life was on the agenda for this OpenAI DevDays, the first of several (the next in London this month and the last in Singapore in November). OpenAI has released a real-time audio API and vision tweaks that allow developers to customize GPT-4o models using images. The company then began model distillation, which takes large-scale AI models like GPT-4o and uses them to fine-tune smaller-scale models.
The focus of this event was not unexpected. OpenAI tempered expectations this summer, saying DevDay would focus on educating developers rather than showcasing products. Nevertheless, what was left out of Tuesday's tight 60-minute keynote raised questions about the progress and status of OpenAI's myriad AI efforts.
We didn't hear anything about a potential successor to OpenAI's nearly year-old image generator, DALL-E 3. We also didn't get any updates on a limited preview of Voice Engine, the company's voice cloning tool. There's no release timeline yet for Sora, OpenAI's video generator, but there is information about Media Manager, an app the company is developing to give creators control over how their content is used in model training. It is full of information.
When asked for comment, an OpenAI spokesperson told TechCrunch that OpenAI is “slowly rolling out.” [Voice Engine] It says it can preview it with more trusted partners and that Media Manager is “still in development.”
But it's clear that OpenAI has reached its limits, and has been for some time.
According to a recent report in the Wall Street Journal, the company's team working on GPT-4o had only nine days to conduct a safety assessment. Fortune reports that many OpenAI staff members thought o1, the company's first “inference” model, wasn't ready for release.
OpenAI, rushing towards a funding round that could raise up to $6.5 billion, is dabbling in a lot of unripe pie. DALL-3 performs worse than image generators such as Flux in many qualitative tests. OpenAI is reportedly improving the model because Sora is very slow in generating footage. And OpenAI continues to postpone the rollout of its GPT Store revenue sharing program, which was originally scheduled for the first quarter of this year.
I'm not surprised that OpenAI is currently suffering from staff burnout and executive departures. If you try to be a jack of all trades, you'll end up being good at nothing and unable to please anyone.
news
AI bill vetoed: California Governor Gavin Newsom has vetoed SB 1047, a high-profile bill that would have regulated AI development in the state. Newsom said in a statement that the bill is “well-intentioned,” but “[not] The best approach to protecting the public from the dangers of AI.”
AI bill passed: Newsom signed other AI regulations into law, including bills that address things like disclosure of AI training data and deepfake nudes.
Y Combinator under fire: Startup accelerator Y Combinator has come under fire for backing AI venture PearAI. PearAI's founders admitted that they essentially cloned an open source project called Continue.
Copilot gets an upgrade: Microsoft's AI-powered Copilot assistant received a makeover on Tuesday. You can now read the screen, think deeply, and speak out loud.
OpenAI co-founder joins Anthropic: One of OpenAI's lesser-known co-founders, Durk Kingma, announced this week that he is joining Anthropic. However, it is unclear what he is working on.
Train AI with customer photos: Meta's AI-powered Ray-Ban has a front-facing camera with various AR features. But it can be a privacy issue. The company hasn't said whether it plans to train the model based on images from users.
Raspberry Pi's AI Camera: Raspberry Pi, which sells small, inexpensive single-board computers, has released the Raspberry Pi AI Camera, an add-on with onboard AI processing.
This week's research paper
AI coding platforms have gained millions of users and raised hundreds of millions of dollars from VC. But do they deliver on their promise to improve productivity?
Probably not, according to a new analysis from engineering analysis firm Uplevel. Uplevel compared data from approximately 800 developer customers. Some of them reported using Copilot, GitHub's AI coding tool, but some did not. Uplevel research shows that developers who rely on Copilot have 41% more bugs and are less susceptible to burnout than those who don't use the tool.
Developers are enthusiastic about AI-powered assisted coding tools despite concerns about security as well as piracy and privacy. The majority of developers responding to GitHub's latest survey said they have adopted some form of AI tools. Companies are also bullish. Microsoft reported in April that Copilot has more than 50,000 business customers.
this week's model
MIT spinoff Liquid AI this week announced its first series of generative AI models, the Liquid Foundation Model (LFM for short).
“So what?” you might ask. The model is the product. New models are released almost every day. LFM uses a new model architecture and achieves competitive scores on various industry benchmarks.
Most models are known as transformers. Transformers, proposed by a team of Google researchers in 2017, have so far become the dominant generative AI model architecture. Transformers power the latest versions of Sora and Stable Diffusion, as well as text generation models such as Anthropic's Claude and Google's Gemini.
However, transformers have their limits. In particular, processing and analyzing large amounts of data is not very efficient.
Liquid claims that its LFM has a reduced memory footprint compared to transformer architectures and can ingest large amounts of data on the same hardware. “Efficiently compressing the input allows LFM to process longer sequences. [of data]” the company wrote in a blog post.
Liquid's LFM is available on many cloud platforms, and the team plans to continue refining the architecture in future releases.
grab bag
If you blinked, you probably missed it. An AI company filed to go public this week.
The San Francisco-based startup, called Cerebras, develops hardware for running and training AI models and competes directly with Nvidia.
So how does Cerebras hope to compete with this large chip company, which controlled 70% to 95% of the AI chip segment as of July? Regarding the performance, Cerebras says: The company claims it has the potential to outperform Nvidia's hardware because its flagship AI chips are sold directly and delivered as a service via the cloud.
However, Cerebras has yet to capitalize on this claimed performance advantage. The company posted a net loss of $66.6 million in the first half of 2024, according to SEC filings. And last year, Cerebras reported a net loss of $127.2 million on revenue of $78.7 million.
According to Bloomberg, Cerebras could aim to raise up to $1 billion through an IPO. The company has raised $715 million in venture capital and was valued at more than $4 billion three years ago.