Welcome to TechCrunch’s AI newsletter!
This week in AI, Gartner released a report predicting that around 30% of generative AI projects in the enterprise will be abandoned after the proof-of-concept phase by the end of 2025. Reasons cited include poor data quality, inadequate risk controls, and escalating infrastructure costs.
One of the major obstacles hindering generative AI adoption is the unclear business value, according to the report. The implementation of generative AI on an organization-wide scale comes with significant costs, ranging from $5 million to $20 million. For example, a simple coding assistant can cost between $100,000 and $200,000 upfront with recurring costs of over $550 per user per year.
Despite the high costs, a recent survey from Upwork revealed that AI has proven to be a burden for many workers rather than enhancing productivity. Nearly half of workers using AI reported having no idea how to achieve the productivity gains expected by their employers, and over three-fourths believe that AI tools have decreased productivity and added to their workload.
While there is robust VC activity in the AI space, anecdotal evidence suggests that generative AI, which has unresolved technical issues, can often be more trouble than it’s worth.
Just recently, Bloomberg published a piece highlighting a Google-powered AI tool used for analyzing patient medical records, which failed to consistently deliver reliable health information in some cases.
Companies are increasingly expecting more from AI technology, and vendors must manage expectations unless there are significant research breakthroughs to address the limitations.
News
SearchGPT: OpenAI announced SearchGPT, a search feature to provide timely answers to questions by drawing information from web sources.
Bing gets more AI: Microsoft introduced Bing generative search, an AI-powered search experience that generates summaries in response to search queries by aggregating information from the web.
X opts users in: X (formerly Twitter) quietly changed its data policy to include user data in training its chatbot Grok, sparking concerns from regulators and users regarding privacy.
EU calls for help with AI: The European Union started a consultation on rules for general-purpose AI models under its AI Act framework.
Perplexity details publisher licensing: AI search engine Perplexity will share advertising revenue with news publishers when their content is surfaced by the chatbot in response to queries.
Meta rolls out AI Studio: Meta launched AI Studio to allow creators to build personalized AI-powered chatbots.
Commerce Department endorses “open” models: The U.S. Commerce Department supported the use of “open-weight” generative AI models like Meta’s Llama 3.1 but recommended new capabilities to monitor potential risks.
$99 Friend: Avi Schiffmann is developing a $99 AI-powered device called Friend, designed to serve as a companion, though its effectiveness remains unclear.
Research paper of the week
OpenAI is exploring rule-based rewards (RBRs) as an alternative to reinforcement learning from human feedback to guide AI models to follow instructions and safety guidelines.
RBRs break down desired behaviors into specific rules to train a reward model, demonstrating better safety performance compared to models trained solely on human feedback.
Model of the week
DeepMind trained two AI systems, AlphaProof and AlphaGeometry 2, to solve complex math problems from the International Mathematical Olympiad, achieving silver medal-level performance on multiple questions.
While these systems showcased impressive reasoning capabilities, their ability to handle open-ended problems with multiple solutions remains limited.
Grab bag
Stability AI released Stable Video 4D, a generative AI model that creates multiple video clips from a single object video, potentially beneficial for game development, video editing, and virtual reality.
The technology is continuously being refined to handle real-world videos beyond its current capabilities.
