**The AGI Bubble is Losing Steam** *Cooling Off* OpenAI's forthcoming large language model, code-named Orion, is reportedly underwhelming, showing less progress than GPT-4 did over its predecessor, GPT-3. Bloomberg reports that some researchers at OpenAI believe there are no improvements, particularly in areas like coding. Furthermore, Google’s new Gemini model is not meeting internal expectations, and the timeline for Anthropic's much-anticipated Claude 3. 5 Opus remains uncertain. These challenges suggest that the current method of enhancing AI models through "scaling" might be reaching its limits. Continuing with these costly developments without significant performance improvements could foreshadow economic difficulties if the goal is achieving artificial general intelligence. According to Margaret Mitchell, chief ethics scientist at the AI startup Hugging Face, "The AGI bubble is bursting a little bit, " indicating that new training approaches may be necessary to reach human-like intelligence and versatility. *Gluttonous Tech* The growth strategy for generative AI has primarily been scaling: increasing the power of models by expanding their size, which involves more processing power from companies like Nvidia and using vast amounts of web-scraped training data. However, as models grow, so do their energy demands. The cost is steep—Microsoft, for instance, is considering reviving nuclear plants to support its AI data centers.

With free training data drying up, tech companies turn to synthetic data but still face difficulties obtaining unique, high-quality datasets without human involvement, as Lila Tretikov, AI strategy head at New Enterprise associates, noted. To illustrate these expenses: Anthropic CEO Dario Amodei mentioned in a podcast, as quoted by Bloomberg, that developing a cutting-edge AI model currently costs about $100 million, with predictions of it rising beyond $10 billion by 2027. *Best Days Behind* This year, Anthropic updated its Claude models, bypassing Opus, and any references to a near-future release were removed from its website. According to Bloomberg, Opus showed only slight improvements relative to its size and cost. Similarly, Google's Gemini software hasn’t met expectations, and major advancements in its large language model are lacking. These obstacles are not unbeatable, but it seems unlikely the AI industry will maintain the rapid advancement pace witnessed over the past decade. "We got very excited for a brief period of very fast progress, " Noah Giansiracusa, a mathematics associate professor at Bentley University, told Bloomberg. "That just wasn't sustainable. " For more on AI: An AI expert warns of an imminent crash as improvements in AI hit a wall.