After a brief break, we're back with some show notes from OpenAI's DevDay.
Yesterday morning's keynote in San Francisco was notable for its subdued tone, a contrast to CEO Sam Altman's vindictive, hype-beast speech last yr. This DevDay, Altman didn't take the stage to unveil shiny recent projects. He didn't even show up; Olivier Godement, Head of Platform Product, moderated.
On the agenda for this primary of several OpenAI DevDays – the following one is in London this month and the last in November in Singapore – quality of life improvements were on the agenda. OpenAI has released a real-time speech API in addition to visual tuning that enables developers to customize its GPT-4o model using images. And the corporate introduced model distillation, which involves using a big AI model like GPT-4o to fine-tune a smaller model.
The narrow focus of the event was not unexpected. OpenAI tempered expectations this summer, saying DevDay would give attention to training developers relatively than showcasing products. Still, what was disregarded of Tuesday's transient, 60-minute keynote raised questions on the progress — and standing — of OpenAI's myriad AI efforts.
We haven't heard what might replace OpenAI's nearly year-old DALL-E 3 image generator, nor have we received an update on the limited preview for Voice Engine, the corporate's voice cloning tool. There's no release schedule yet for Sora, OpenAI's video generator, and Mama has the word on Media Manager, the app the corporate says it's developing to present developers control over how their content looks in model training be used.
When reached for comment, an OpenAI spokesperson told TechCrunch that OpenAI is “slowly rolling out the (Voice Engine) preview to more trusted partners” and that Media Manager is “still in development.”
But it seems clear that OpenAI is at capability – and has been for a while.
According to current status reporting As the Wall Street Journal reported, the corporate's teams working on GPT-4o had just nine days to conduct security assessments. Assets Reports that many OpenAI employees thought that o1, the corporate's first “reasoning” model, was not yet ready for release.
As it heads toward a funding round that would raise as much as $6.5 billion, OpenAI has its fingers in quite a lot of unbaked cakes. In many qualitative tests, DALL-3 is worse than image generators like Flux. Sora is allegedly Footage generation is so slow that OpenAI is overhauling the model. and OpenAI continues to delay the launch of the revenue share program for its bot marketplace, the GPT Store, which was originally scheduled for the primary quarter of this yr.
I'm not surprised that OpenAI is facing this now Staff burnout and departures of executives. If you are trying to be a jack of all trades, you'll find yourself being a master of none – and pleasing nobody.
News
AI bill vetoed: California Governor Gavin Newsom vetoed SB 1047, a high-profile bill that may have regulated AI development within the state. In an announcement, Newsom called the bill “well-intentioned” but “(not) one of the best approach” to guard the general public from the risks of AI.
Passed AI laws: Newsom has signed other AI regulations into law — including bills addressing disclosure of AI training data, deepfake nudes and more.
Y Combinator criticized: Startup accelerator Y Combinator is under fire after backing an AI company, PearAI, whose founders admitted they essentially cloned an open source project called Continue.
Copilot is updated: Microsoft's AI-powered Copilot assistant got a makeover on Tuesday. It can now read your screen, think deeply, and check with you out loud, amongst other things.
OpenAI co-founder joins Anthropic: Durk Kingma, considered one of OpenAI's lesser-known co-founders, this week announced he’ll join Anthropic. However, it’s unclear what he will likely be working on.
AI training based on customer photos: Meta's AI-powered Ray-Bans feature a camera on the front for various AR functions. But it could develop into a privacy issue — the corporate won't say whether it plans to coach models using images from users.
Raspberry Pi AI camera: Raspberry Pi, the corporate that sells tiny, inexpensive single-board computers, has released the Raspberry Pi AI Camera, an add-on with built-in AI processing.
Research paper of the week
AI coding platforms have acquired hundreds of thousands of users and attracted a whole bunch of hundreds of thousands of dollars from VCs. But do they deliver on their guarantees to extend productivity?
Maybe not, so a recent evaluation from Uplevel, a technical evaluation company. Uplevel compared data from about 800 of its developer customers – a few of whom said they used GitHub's AI coding tool Copilot, a few of whom didn't. Uplevel found that developers who relied on Copilot made 41% more errors and were no less susceptible to burnout than those that didn't use the tool.
Developers are showing enthusiasm for AI-powered assistive coding tools despite concerns not only about security but additionally copyright infringement and privacy. The overwhelming majority of developers who participated in GitHub's latest survey said they’ve adopted AI tools in some form. Companies are also optimistic – Microsoft reported in April that Copilot had done so over 50,000 corporate customers.
Model of the week
Liquid AI, a spin-off from MIT, announced this week its first series of generative AI models: Liquid Foundation Models, or LFMs.
“So what?” one might ask. Models are a commodity – recent ones come onto the market practically on daily basis. Well, LFMs use a novel model architecture and achieve excellent competitive results on quite a few industry benchmarks.
Most models are so-called transformers. The Transformer was proposed by a team of Google researchers in 2017 and has change into the dominant generative AI model architecture. Transformers are the idea for Sora and the newest version of Stable Diffusion, in addition to text-generating models like Anthropic's Claude and Google's Gemini.
But transformers have limitations. In particular, they are usually not very efficient at processing and analyzing large amounts of information.
Liquid claims that its LFMs have lower memory footprints in comparison with Transformer architectures, allowing them to accommodate larger amounts of information on the identical hardware. “By efficiently compressing inputs, LFMs can process longer sequences (of information),” the corporate wrote in a Blog post.
Liquid's LFMs can be found on quite a few cloud platforms and the team plans to further refine the architecture with future releases.
Lucky bag
If you blinked, you almost certainly 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 plan to compete against the chip giant? ordered Between 70% and 95% of the AI ​​chip segment as of July? On performance, says Cerebras. The company claims that its flagship AI chip, which it sells each directly and offers as a service through its cloud, can outperform Nvidia's hardware.
But Cerebras has yet to convert this alleged performance advantage into profit. The company reported a net lack of $66.6 million in the primary half of 2024. per filing with the SEC. And for last yr, Cerebras reported a net lack of $127.2 million on revenue of $78.7 million.
Cerebras could seek to lift as much as $1 billion through its IPO. after to Bloomberg. To date, the corporate has raised $715 million in enterprise capital and was valued at over $4 billion three years ago.
