Sara Altman
Simon Couch
Roundups

2025-10-24 AI Newsletter

2025-10-24
A collection of hexagons, each containing a different icon. On the left are three outlined heads: two with glasses, one a robot. The hexagons include a mall, a bridge, a raccoon, a Viking, and a goose.

External news

 

Claude Haiku 4.5

 

Last week, Anthropic released Claude Haiku 4.5, a cheaper, smaller Claude model. On Anthropic’s own software engineering benchmark, Haiku 4.5 performed as well as the more expensive Sonnet 4. Our sense is that Haiku 4.5 probably is as good as Sonnet 4 (but not Sonnet 4.5), and so if you’re looking for a cheaper model, Haiku 4.5 is a good option. However, Haiku 4.5 is still more expensive than other cheaper models from the major labs, and none of Anthropic’s models really qualify as inexpensive. 

 

Claude Skills

 

The same day, Anthropic also announced Claude Skills. Anthropic defines skills as, “folders that include instructions, scripts, and resources that Claude can load when needed.” For example, you might create a documentation skill with your organization’s style guide, templates, and example docs, helping Claude produce documentation that fits your requirements and style.

Skills are useful because they give the model additional capabilities, but the model only loads the required information when it’s about to use that skill. This contrasts with MCP tools and prompts, which always live in the model’s context and can cause significant context bloat. We at Posit have already been getting a lot of use out of them! 

 

AI water and energy use

 

Andy Masley’s recent blog post takes a detailed look at the water use of AI. He argues that AI’s water use has been dramatically overstated and is not a meaningful environmental issue. He estimates that all U.S. data centers consume about 0.2% of the nation’s freshwater, with AI-specific water usage around 0.008%, comparable to the water needs of a few small towns. 

In contrast, the electricity usage of data centers that serve AI models may deserve some concern, as outlined in this post from Hannah Ritchie. From around 2005 until recently, power demand and generation were relatively flat, but both are now beginning to rise. When demand grows faster than generation in a specific area, electricity prices tend to increase. We’ve already seen this in areas where new data centers have added significant load to local grids. 

 

Posit news

 

As of version 2025.10.0-199, Positron now supports GitHub Copilot as a model provider for chat in Positron Assistant. Previously, GitHub Copilot was only available for inline code completion (and is still the only option for code completion). Now, if you have GitHub Copilot added as a model provider, its models, chat participants, and tools are available for sidebar chat, inline chat, and inline code completion.

 

Terms

 

Artificial general intelligence (AGI) refers to AI systems with intelligence that matches or exceeds human intelligence. Since the arrival of LLMs, speculation about AGI has intensified: when will it arrive, is it possible, does it matter, and could it be catastrophic? The AI 2027 group, which includes Daniel Kokotajlo, a former OpenAI researcher who has written extensively about the potential risks of highly intelligent AI since leaving the company last year, believes AGI is likely to emerge relatively soon. Many others remain skeptical.

This week, in an interview with Dwarkesh Patel, Andrej Karpathy, cofounder of OpenAI and a leading figure in AI, suggested that AGI is still at least a decade away. He described our current period as the “decade of agents,” arguing that while today’s agentic systems are impressive and increasingly useful, they still lack the core capabilities needed to truly function like humans performing cognitive tasks. You can watch the full (2.5-hour) interview and read the transcript here. 

 

Learn more

 

  • Jack Clark shared the full transcript of his talk at The Curve, a conference on “AI’s biggest questions.” The talk is a grounded reflection on Clark’s two decades of experience in AI research. 
  • A recent preprint finds that state-of-the-art AI models are highly sycophantic, affirming “users’ actions 50% more than humans do, and [doing] so even in cases where user queries mention manipulation, deception, or other relational harms.” It also finds that people tend to prefer sycophantic responses and that these responses reinforce problematic beliefs.
  • Andrej Karpathy released nanochat, a complete, open-source implementation of a ChatGPT-like system that can be trained for about $100 in compute. This was one of the topics discussed in his interview with Dwarkesh Patel, mentioned above. 
Tags: AI AI Newsletter