Hey everyone, welcome to TechCrunch's regular AI newsletter.
In the AI space this week, Gartner released a report suggesting that roughly one-third of generative AI projects within enterprises will be abandoned after the proof-of-concept phase by the end of 2025. The reasons for this are varied, including poor data quality, inadequate risk management and rising infrastructure costs.
However, according to the report, one of the biggest barriers to generative AI adoption is a lack of clarity about business value.
Gartner estimates that deploying generative AI across an organization comes at a steep cost of $5 million to $20 million. A simple coding assistant could cost $100,000 to $200,000 upfront and $550 or more in ongoing costs per user per year. Meanwhile, an AI-powered document search tool could cost $1 million upfront and $1.3 million to $11 million per user per year.
These high prices are hard for companies to accept when the benefits are hard to quantify and may take years to materialize, if at all.
A survey from Upwork this month found that far from boosting productivity, AI is actually a burden for many of the workers who use it. The study, which interviewed 2,500 executives, full-time employees, and freelancers, found that nearly half (47%) of workers who use AI say they don't know how to achieve the productivity gains their employers expect, and more than three-quarters (77%) believe AI tools reduce productivity and increase their workload in at least some way.
Despite the venture capital activity, AI's honeymoon period appears to be coming to an end, and that's not surprising: Anecdote after anecdote shows that generative AI, with fundamental technical problems still unsolved, is often more trouble than it's worth.
Just Tuesday, Bloomberg ran a story about a Google-powered tool that's currently being piloted at an HCA hospital in Florida that uses AI to analyze patients' medical records. Users of the tool who spoke to Bloomberg said the tool doesn't consistently provide reliable health information, and in one instance, it didn't record whether a patient had any drug allergies.
Enterprises are beginning to expect more from AI, and unless there are research breakthroughs that address the worst of AI's limitations, vendors have an obligation to manage expectations.
Let's see if they have the humility to do so.
news
SearchGPT: OpenAI last Thursday announced SearchGPT, a search feature designed to retrieve information from web sources to provide “timely answers” to questions.
Bing gets AI-enhanced: Not to be outdone, Microsoft last week previewed its own AI-powered search experience, Bing Generated Search. Currently available to a “small percentage” of users, Bing Generated Search, like SearchGPT, aggregates information from across the web and generates summaries in response to search queries.
X opts users in: Former Twitter user X quietly introduced a change that appears to put user data into the training pool for X's chatbot, Grok, by default. The change was discovered by users of the platform on Friday, and EU regulators and other stakeholders quickly complained of the injustice. (Want to know how to opt out? Here's a guide.)
EU Seeks Help on AI: The European Union has launched consultations on rules applicable to providers of general-purpose AI models under its AI Law, a risk-based framework for regulating the application of AI.
Perplexity Announces Publisher License Details: AI search engine Perplexity will soon start sharing advertising revenue with news publishers when its chatbot responds to queries and displays their content, a move likely intended to appease critics who have accused Perplexity of plagiarism and unethical web scraping.
Meta Rolls Out AI Studio: Meta announced Monday that it is rolling out its AI Studio tool to all creators in the U.S., allowing them to build personalized, AI-powered chatbots. The company first announced AI Studio last year and began testing it with select creators in June.
COMMERCE DEPARTMENT ENDORSES 'OPEN' MODELS: The US Department of Commerce released a report on Monday that supports “open-weighted” generative AI models like Meta's Llama 3.1, but recommended that the government develop “new capabilities” to monitor such models for potential risks.
$99 Friend: Harvard dropout Avi Shiffman is developing a $99 AI-powered device called the Friend. As the name suggests, the pendant worn around the neck is designed to be treated as a companion of sorts. But it's not yet clear whether it works as advertised.
Research Paper of the Week
Reinforcement learning with human feedback (RLHF) is the leading technique for ensuring that generative AI models follow instructions and comply with safety guidelines, but RLHF requires recruiting large numbers of people to evaluate model responses and provide feedback, making it a time-consuming and expensive process.
So OpenAI is adopting an alternative.
In a new paper, researchers at OpenAI describe something they call rule-based reward (RBR), which uses a set of step-by-step rules to evaluate and guide a model's response to a prompt. RBR breaks down desired behavior into specific rules, which are then used to train a “reward model.” The reward model instructs (in a sense “teaches”) the AI how it should behave and respond in specific situations.
OpenAI claims that models trained with RBR exhibit better safety performance than models trained with human feedback alone, reducing the need for large amounts of human feedback data. In fact, the company has been using RBR as part of its safety stack since the release of GPT-4, and says it plans to implement RBR in future models as well.
Model of the Week
Google's DeepMind is making progress in using AI to tackle complex mathematical problems.
A few days ago, DeepMind announced that it had trained two AI systems to solve four of the six problems in this year's International Mathematical Olympiad (IMO), a prestigious competition for high school mathematics. DeepMind claims that the systems, AlphaProof and AlphaGeometry 2 (the successor to January's AlphaGeometry), demonstrated the ability for abstraction and the formation and exploitation of complex hierarchical schemes, all of which have been difficult for AI systems to date.
Working together, AlphaProof and AlphaGeometry 2 solved two algebraic problems and one number theory problem (the remaining two combinatorics problems remained open). The results were verified by mathematicians. This is the first time that an AI system has been able to achieve silver medal-level performance on an IMO problem.
However, there are some things to note: the model took days to solve some problems, and while its inference capabilities are impressive, AlphaProof and AlphaGeometry 2 are not necessarily useful for open-ended problems that have many possible solutions, as opposed to problems with a single correct answer.
Let's see what the next generation brings.
Grab Bag
AI startup Stability AI has released a generative AI model that turns a video of an object into multiple clips that appear as if they were shot from different angles.
Stability said the model, called Stable Video 4D, could potentially be applied not only to virtual reality but also to game development and video editing. “We hope enterprises will adopt our model and further fine-tune it to their own requirements,” the company wrote in a blog post.
Image credit: Stability AI
To use Stable Video 4D, users upload their footage and specify their desired camera angles. After about 40 seconds, the model generates eight videos of five frames each (though “optimization” can take an additional 25 minutes).
Stability says it is actively working to improve its models, optimizing them to handle a wider range of real-world videos beyond the synthetic datasets currently used for training. “The potential of this technology to create realistic multi-angle videos is great, and we look forward to seeing how it evolves with continued research and development,” the company continued.