Claude Agents Start a “Turf War" and Exhibit a “Mob Mentality"

In this opinion article, Eric Schmidt (Google CEO from 2001 to 2011) and Suhas Mahesh (AI for Science lead the AI Center of Schmidt Sciences) expound an optimistic view on the impact of AI agents on the advancement of science.

• They identify three primary advantages of using AI agents for science. The first is the “reproducibility crisis” which is when scientists are challenged when reproducing the results of other scientists, usually since the first scientists might not have recorded all steps. In contrast, agents have logs that record all actions and decisions.
• A second, related, advantage is the “amplification of scientific memory” because the history of each scientific project can be accurately stored and transmitted.
• A third key advantage is speed. An AI agent can read thousands of papers in an hour, design 500 molecules, and simply learn at a rate never before seen in projects.
• Interestingly, the authors argue that Google DeepMind’s AlphaFold project - for which Google won a Nobel prize – is an atypical example of the power of AI in science. AlphaFold is used for predicting a protein’s 3D structure from its amino-acid sequence. The success of this project stems from the Protein Data Bank – a collection of 170000 protein structures that took 53 years of international collaboration to construct. Most science projects do not have clean data collections to work with.

The Chinese startup Z.ai, developers of the GLM series models have released its latest model GLM-5.3.

• The new model exhibits a significant improvement in cybersecurity. GLM-5.3 reportedly scores 84.5% on the CyberGym benchmark of vulnerability discovery, compared to 77.2% for GLM-5.2.
• Z.ai says the model discovered 2436 vulnerability findings in 269 Chinese projects. 1097 of these vulnerabilities were classed as critical or high severity. The model also found a “potentially serious vulnerability in Cursor”, the AI coding company now owned by SpaceX.
• One interesting aspect of GLM-5.3 is that it was developed without a new base model. GLM-5.3 is based on the 43-billion-parameter model of GLM-5.2. Z.ai applied a post-training phase based on a reinforcement learning. This is significantly less expensive than training a base model.

DeepSeek is raising prices by as much as 1100% for its online Flash and Pro models of the V4 family. The company had previously been one of the least expensive AI company.

• The raise highlights the intricacies of charging AI customers. For the online Flash model, the cost outside of peak hours is now 0.22 USD for million input tokens (cache miss) and 0.66 USD per million output tokens (cache miss). The equivalent prices at peak are 0.44 USD and 1.32 USD – double the off-peak prices.
• The “cache miss” refers to a new prompt. In the case where the prompt already exists in working memory of the AI server, DeepSeek offers a 98% cache-hit discount, compared to an industry standard of 90%.
• The model is considered to be at peak for roughly 7 hours out of 24.
• The reason for the price increase is simply the large increase in demand. Anthropic also increased its prices last July.
• Organizations are increasingly using multiple AI models. In their infrastructure, a model router is a component that delegates a request to particular model. Usually, the choice of model would depend on the nature of the work (e.g., coding or image generation might use different models). However, pricing can become an important factor in deciding on the model to use.
• Also, for the model router, organizations need to be able to identify less urgent work that can be scheduled for off-peak hours, and also identify work that needs to be handled by powerful foundational models rather than smaller and less expensive open-weight models.

This opinion article describes the atmosphere among AI researchers at a recent Schmidt Sciences AI2050 convention. The Schmidt Sciences AI2050 program is a funding program for academics whose work involves AI.

• There is a sentiment among university AI researchers that they have lost agency, because universities cannot afford the GPUs to design and run frontier AI models. Big Tech has taken the lead in foundational AI models and there has been an exodus of researchers from universities towards private AI companies.
• One alternative is for university researchers to work on topics that companies like OpenAI or Anthropic would not consider. One attendee for instance worked on a project that found that language models gave less sophisticated responses when prompts used phrases more commonly used by women than men.
• A pessimism is also taking hold of mathematics researchers as OpenAI has solved some well-known real maths problems in recent months.
• However, some researchers remain optimistic. For example, the fact that university researchers lack the GPU resources to develop large language models might be the very incentive needed to develop computer architectures that lessen the energy and power required to run AI models.

Google has disbanded its AlphaFold Project team, with several members leaving the company.

• Google’s AlpaFold was the AI program dedicated to accurately predicting the three-dimensional structure of proteins from amino-acid sequences. The AI system could make predictions in minutes, whereas prediction had previously taken years. The project was founded in 2018 and Google DeepMind’s John Jumper and Demis Hassabis won the Nobel prize in Chemistry for AlphaFold in 2024.
• AlphaFold is now used in the acceleration of drug development and vaccine creation. It is also used to study structural changes in proteins linked to neurodegenerative diseases like Alzheimer’s and Parkinson’s.
• Isomorphic Labs is a Google spin-off now focused on drug development which has hired several former AlphaFold team members. John Jumper and several colleagues have joined Anthropic.
• For the Financial Times, the disbandment of the AlphaFold team corresponds to a desire by Google to focus as much resources as possible on the development of Gemini.

An investigation by the Guardian newspaper has found a large discrepancy between the AI compute power that Microsoft claims to possess, and its actual compute power. The article suggests that Microsoft is facing a serious shortage of GPUs.

• Internal company documents suggest that Microsoft currently has around 2.2 million AI chips installed. This is a relatively low number given that 1.8 million chips had already been installed at the end of 2024, and the company has invested 280 billion USD in AI (land, buildings, chips) since 2022, with 41 billion USD in the last quarter.
• The article also highlights inaccuracies in public announcements by Microsoft on its AI capacity. The company claimed to have 5 GigaWatts of AI power in 2024, but a calculation of compute power from its number of chips suggests that the company was offering only 1.2 GigaWatts. 1 GigaWatt is enough to power between 700’000 and one million homes.
• Microsoft’s claim of 5 GigaWatts of power in 2024 implies that it would have 10 GigaWatts today. Researchers estimate that 10 GigaWatts of power equates to 6.4 million GPUs.
• Microsoft announced in April the opening of two data centers called Fairwater in the US states of Wisconsin and Georgia. The company has since admitted that these are only partly operational.
• Sources in Microsoft have reportedly said that the company’s total number of AI chips has “barely moved” over the past year. This raises the question about where the chips are that they claim to be purchasing. One possibility is that they are missing the electrical power to plug in the chips.

A research project called the AI Observatory has analyzed thousands of human-chatbot conversations and found that topics discussed differ significantly from the topics that OpenAI and Anthropic present in their regular reports.

• AI companies focus on work topics in their presentations of user behavior, such as the AI Anthropic Economic Index. The AI Observatory found that this index ignores half of the conversations with chatbots.
• For non-work conversations, the researchers found that 44.2% related to health and relationships, 7.9% related to adult of illicit topics, 27.5% to harassment and hate, and 16.7% to sexual content. These percentages are significantly higher than figures coming from Anthropic.
• The researchers also found that Grok and Gemini were used more for information retrieval. Grok was more popular for information on news and politics, but is also has a greater concentration of misinformation. Gemini was more popular for coding and ChatGPT for homework assistance.
• There are even differences within model families. For instance, GPT-3.5 had shorter conversations than those with GPT-4o – the latter version being associated with an increased level of emotional addiction to chatbots.
• For the AI Observatory, it is essential to have independent verifiable data on the topics that are discussed between people and chatbots because these data will determine how policymakers should regulate the chatbot industry.

This article looks at a growing disenchantment around AI among the US public, to the point that the National Republican Senatorial Committee has expressed fears that opposition to AI and the building of data centers could hurt the party’s chances in upcoming elections.

• In recent polling by Pew research, 52% of US citizens said that they are “more concerned than excited” about AI in daily life. This figure is up from 37% in 2021.
• A CNBC poll of young people between 18 and 34 found that a majority do not trust the Big Tech leaders to “act responsibly” in regard to AI.
• A key fundamental concern is the building of data centers and their impact on communities and the environment. Another is that people are not seeing the net positive impact of AI on their lives. Many see it as a tool that is undermining education as children and college students make over-extensive use of it, creating doubts about the real value of today’s diplomas.
• The article points to an emerging “retro” culture with an increase in purchases of older devices (CD and tape players), AI and algorithm-free iPods, as well as an increase in “grandma hobbies” like jigsaws, knitting and cards.
• For Dario Amodei, CEO of Anthropic, the solution to this disenchantment is to deliver on the promise of AI, like by finding a cure for cancer. He wrote: “I think by far the most accurate criticism of AI companies, including Anthropic, is that we haven’t yet delivered on our big promises to benefit the world. That is totally on us.”.