Hello everyone and welcome to TechCrunch's regular AI newsletter. If you'd like to receive this newsletter in your inbox every Wednesday, sign up here.
This week in AI, if a story in The Information is to be believed, OpenAI's next major product announcement is on the horizon.
The Information reported on Tuesday that OpenAI plans to release Strawberry, an AI model that can effectively fact-check itself, within the next two weeks. Strawberry will be a standalone product but will also be integrated with OpenAI's AI-powered chatbot platform, ChatGPT.
Strawberry is reported to be better at programming and math problems than other top generative AI models (including OpenAI's own GPT-4o), and it avoids some of the inference pitfalls that typically plague those models. But the improvement comes at a price: Strawberry is said to be slow. Really slow. Sources say the model takes 10 to 20 seconds to answer a single question.
Indeed, OpenAI is likely to position Strawberry as a model for mission-critical tasks where accuracy is paramount, which could be embraced by many companies frustrated by the limitations of today's generative AI technology: A survey this week by HR specialist Peninsula found that inaccuracy is a main concern for 41% of companies considering deploying generative AI, and Gartner predicts that a third of all generative AI projects will be abandoned by the end of the year due to barriers to adoption.
But while some businesses may not care about chatbot latency, the average person probably does.
Hallucinatory tendencies aside, today's models are fast. Incredibly fast. We've grown accustomed to them. In fact, this speed makes interactions feel more natural. If Strawberry's “processing time” were orders of magnitude longer than existing models, it would be hard to avoid the perception that Strawberry is in some way a step backwards.
This assumes the best-case scenario, that is, that Strawberry answers questions correctly consistently. If it's still error-prone, as reports suggest, the long wait times would be even more unacceptable.
OpenAI is no doubt feeling the pressure to deliver, as it spends billions on AI training and talent development. The company's investors and potential new backers will likely be hoping to see profits sooner rather than later. But rushing to release an unfinished model like Strawberry and then considering charging a significantly higher price for it seems unwise.
I think the smarter move would be to let the technology mature a bit more. As the race for generative AI heats up, OpenAI may not have the luxury.
news
Apple Rolls Out Visual Search: The new camera control button on the iPhone 16 and 16 Plus can activate a feature Apple calls “visual intelligence,” which is essentially a reverse image search combined with text recognition. The company is partnering with third parties, including Google, to power its search results.
Apple isn't dabbling with AI: Devin writes that many of Apple's AI-generated features are ultimately pretty basic, contrary to what the company's hyperbolic marketing would have you believe.
Audible Trains AI for Audiobooks: Amazon's audiobook business, Audible, announced that it will create new audiobook recordings using AI trained on the voices of professional narrators. Narrators will be compensated through a share of royalties per title for audiobooks created using the AI voices.
Musk denies Tesla deal with xAI: Elon Musk disputed a Wall Street Journal report that he was in revenue-sharing talks with another company, xAI, to allow one of his companies, Tesla, to use its generative AI models.
Bing introduces deepfake removal tool: Microsoft announced that it is working with StopNCII, an organization that helps victims of revenge porn create digital fingerprints of explicit images, whether they are real or not, to help remove non-consensual pornography from Bing search results.
Google's Ask Photos Launches: Google's AI-powered search feature, Ask Photos, began rolling out last weekend to some Google Photos users in the U.S. With Ask Photos, you can ask complex questions like, “Show me the best photos of the national parks I've visited,” “What did you order last time at this restaurant?” or “Where did you camp last August?”
US and EU Sign AI Treaty: At their summit last week, the US, UK and EU signed a treaty on AI safety developed by the Council of Europe (COE), an international standards and human rights body. The COE describes the treaty as “the first-ever legally binding international treaty aimed at ensuring that the use of AI systems is fully compatible with human rights, democracy and the rule of law.”
Research Paper of the Week
All biological processes depend on protein-protein interactions, which occur when proteins bind to each other. Proteins that bind to specific target molecules, or “binder” proteins, are used in drug development and disease diagnosis.
However, creating binder proteins is often a laborious, expensive process that carries with it the risk of failure.
Seeking an AI-powered solution, Google's AI lab DeepMind developed AlphaProteo, a model that predicts protein binding to a target molecule. Given a few parameters, AlphaProteo can output candidate proteins that bind to a molecule at a specified binding site.
In tests with seven target molecules, AlphaProteo generated protein binders with “binding affinities” (molecular binding strength) that were 3 to 300 times better than those that could be produced by traditional binder discovery methods. Additionally, AlphaProteo was the first to successfully develop binders for a protein (VEGF-A) that is associated with complications resulting from cancer and diabetes.
But DeepMind acknowledged that AlphaProteo failed on its eighth test, saying strong binding is usually only the first step in creating a protein that could have practical applications.
Model of the Week
There are new, incredibly powerful generative AI models out there that anyone can download, tweak, and run.
The Allen Institute for Artificial Intelligence (AI2), in collaboration with startup Context AI, has developed an English text generation model called OLMoE, which has a mixture of experts (MoE) architecture with 7 billion parameters. (A “parameter” roughly corresponds to the model's problem-solving ability; models with more parameters generally perform better than models with fewer parameters, but not always.)
MoE breaks down data processing tasks into subtasks and delegates them to smaller, more specialized “expert” models. This is not new, but what makes OLMoE notable is the fact that it is openly licensed, plus outperforms many models in its class across a range of applications and benchmarks, including Meta's Llama 2, Google's Gemma 2, and Mistral's Mistral 7B.
Several variations of OLMoE, as well as the data and code used to create them, are available on GitHub.
Grab Bag
It was Apple week this week. The company held an event on Monday where it unveiled new iPhones, Apple Watch models, and apps. If you weren't able to tune in, here's a recap.
Apple Intelligence, Apple's suite of AI-powered services, predictably got the most attention, with Apple reaffirming that ChatGPT will be integrated into the experience in some key ways, but curiously there was no mention of any AI partnerships beyond its previously announced deal with OpenAI, despite Apple having lightly hinted at such a partnership earlier this summer.
At WWDC 24 in June, senior vice president Craig Federighi confirmed Apple's plans to work with additional third-party models in the future, including Google's Gemini. “We don't have anything to announce at this point, but that's the broad direction we're heading,” Federighi said.
He has not been heard from since.
Perhaps gathering the necessary paperwork is taking longer than expected, there were technical issues, or perhaps Apple's potential investment in OpenAI has antagonized some of the model partners.
Either way, ChatGPT will likely be the only third-party model for Apple Intelligence for the foreseeable future. Sorry, Gemini fans.