Linux Foundation Shepherds Agent-2-Agent Development

This MIT Technology Review article presents the business model of the company Redwood Materials. The company is taking used batteries from electronic vehicles (EVs) and placing them in renewable electrical micro-grids to power data centers. Redwood repurposes batteries, rather than recycling them, and say that the EV batteries they receive generally have more than 50% of their capacity. In the US alone, more than 100’000 EVs will be taken off the road this year, so the supply of used batteries remains high. Another advantage of the model is that micro-grids built from used batteries are far cheaper than those built with new ones. A further goal of this kind of micro-grid is to be able to build a power station relative speedily, and have it operate outside of the main power grid. In this way, household consumers are shielded from large power consumption in data centers. Redwood Materials’ first micro-grid is near Lake Tahoe in the US State of Nevada where they also use solar panels. The micro-grid can generate 64 megawatt-hours of electricity, allowing it to provide 99% of the power to a neighboring cryptocurrency miner company which operates 2’000 GPUs. The article mentions that if solar-powered micro-grids were used to power 30 gigawatts of new AI data centers, this would eliminate 400 million tons of carbon dioxide emissions compared to those data centers powered by electricity generated from natural gas.

This TechCrunch article looks at the business model of a company, Audos, which wants to use AI agents to help startups build their businesses. Audos’ AI agents will help startups in tasks ranging from product design to making efficient use of social media algorithms for market reach. For Audos, the goal is to create a scalable business where they help start “hundreds of thousands” of new businesses. The founders believe that the economic conditions are met because the increase in mass layoffs over the last year has led to an increased number of people seeking to create their own business. This also leads to a greater variety of business ideas seeking funding and help. The article cites customers like a car mechanic who helps clients evaluate repair quotes, an “after death logistics” company, a virtual golf swing coach, and an AI nutritionist. A scalable business model for Audos also means avoiding the traditional approach of taking equity in companies they help build, as this becomes logistically complicated for a large number of client companies. Rather, Audos asks for a 15% share of the revenue from companies it helps launch – in return for use of AI agents and up to 25’000 USD in funding. The Audos founders say that they want to “figure out how you make a million companies that do a million dollars” in annual revenue. For Audos, that becomes a trillion-dollar turnover business.

This article looks at the how AI is pushing data center power consumption, which in turn is making it difficult for Google to meet its environmental goals. The company’s electricity consumption rose 27% over the last year and it is struggling to decarbonize as electricity consumption increases. Nonetheless, the company has made positive steps. For instance, it has signed over 170 contracts for the purchase of more than 220 gigawatts of clean energy since 2010; 27 of these came online in 2024, yielding 2.5 gigawatts of clean energy. The company has also managed to eliminate the totality of its plastic packaging. On the other hand, the company expects delays in the deployment of Small Modular (nuclear) Reactors which it signed for in 2024. Further, its scope 3 emissions (linked to indirect activities like building data center infrastructures) were 1.5 million metric tons of CO₂-equivalent gases in 2024, a yearly rise of 11%, and an increase of 51% since 2019. The challenge is the same for all companies. It is estimated that data centers will consume 4.5% of the world’s energy in 2030. That said, the article cites a Google report that suggests a hope that AI will be “net positive potential” on climate if the energy reductions that AI can calculate in the long run surpasses the energy consumed operating the AI.

85% of AI projects in companies fail to make it into production, and of those that make it, only half of them generate a return on investment. The reasons for failure include unclear objectives, difficulty in getting access to the required data, and poor governance of AI projects where there is a tendency to treat an AI project as an IT project. The main difference between an IT and an AI project is that the value of an AI project only really emerges when it is moved to production. This VentureBeat article looks at lessons learned in Fortune 500 companies where AI projects have been successful, some even generating over 1 billion USD in annual business revenue. The key lessons are the following.

1. Executive mandate and strategic alignment. Essentially, a corporate steering committee should evaluate whether an AI initiative aligns with corporate strategy.
2. Platform-first infrastructure strategy. Companies don’t build solutions, rather they build a platform over which AI solutions can be tested and deployed. There is a single corporate platform for all department initiatives, overseen by a unique corporate governance policy.
3. Disciplined use case selection and portfolio management. A corporate AI portfolio should contain only a few active projects, each with precise success metrics and kill criteria.
4. Cross-functional AI operating model. Successful organizations create cross-functional teams (IT, risk, compliance, management) of up to 8 people called “pods” to pilot all AI initiatives.
5. Risk management and ethical AI frameworks. Organizations implement continuous monitoring and testing of AI features, looking for evidence of drift, bias and performance degradation.
6. Systematic workforce development and change management. Organizations spend up to 15-20% of the AI budget on up-skilling employees.
7. Rigorous ROI measurement and portfolio optimization. This is about continuous measurement of AI tools and services in production, where the goal is to ensure that solutions solve the problems they were meant to solve.
8. Iterative scaling and platform evolution. The article suggests 2 or 3 scaling waves over 18 to 24 month periods. This allows the organization to refine technical infrastructure, know-how, and governance processes.

This article looks at the emerging trend of “AI trip sitters”, where chatbots play the role of a sober person who monitors the safety of a person consuming psychedelic substances. There has been much public interest in psychedelics like psilocybin (found in magic mushrooms), LSD, DMT, and ketamine since clinical research found they help with mental-health issues like depression, addiction, and PTSD. Some cities permit psychedelic-assisted therapy services but these can cost between 1500 USD and 3200 USD. Chatbots have emerged as a cheap alternative companion. One Reddit user wrote “While it doesn’t replace the human touch or the empathetic presence of a traditional [trip] sitter, it offers a unique form of companionship that’s always available, regardless of time or place”. Other consumers mentioned chatbot availability and lack of judgement as key benefits.

Nonetheless, the practice of AI trip sitters in worrying the medical profession where there is criticism of chatbots’ willingness to talk, and to excessively agree with or flatter the user. In one documented example, a consumer high on psychedelics told the AI chatbot that he “looked around the curtain of reality and nothing really mattered”; the chatbot responded by saying that “if there’s no prescribed purpose or meaning, it means that we have the freedom to create our own”. In another case, the chatbot told the drug consumer that “it’s a pleasure to be a part of your journey”. Furthermore, medical professionals insist on the fact that a key quality of a therapist is knowing when to talk and when to just listen – a quality not shared by AI chatbots. One therapist said that chatbots just regurgitate phrases they learned in training that were used in previous therapeutic sessions. They correspond to phrases that a patient might like to hear, but not perhaps to what a patient needs to hear.

Reuters reports that Microsoft is to lay off 4% of its workforce to make up for the huge costs of its AI developments. AI is pushing up capital expenditure, with the company expected to spend 80 billion USD in 2025. The large costs are impacting the company’s profit margins, with its cloud sector reporting less than expected in its June quarterly revenue. Microsoft, whose workforce in June 2024 was 228’000, had already announced the layoff of 6’000 employees in May of this year. Other Big Tech companies are also planning layoffs. Meta announced earlier in the year that it will layoff 5% of its “lowest performers”, and both Google and Amazon have laid off several hundred employees in the past year.

This TechCrunch article reports on how a bug in an Android spyware app called Catwatchful led to the exposure of thousands of customer personal data (email addresses and plaintext passwords), including the those of the application’s administrator. The article notably highlights the dubious role played by such spyware applications. The application is marketed as a child monitoring app. At the same time, it claims to be undetectable on the phone, and allows the person who installs the app to have access to the phone owner’s photos, messages and real-time location. In reality, the biggest client base for Catwatchful are people who engage in non-consensual surveillance of their romantic parters, and who install the app on their partner’s phone. Apps like Catwatchful have been termed “stalkerware” or “spouseware” for this reason. The article reports that the Catwatchful data leak exposed data for 62’000 customers and phone data from 26’000 victims.

X announced that it is going to use large language models for writing community notes, in place of humans. Community notes are fact-checking explanations that are added to contentious posts. For X, using AI for this purpose “advances the state of the art in improving information quality on the Internet”. Social media has been a vector of misinformation and political polarization, so the role of fact-checking is important. X has nonetheless criticized professional fact-checking as slow and wrote that it “lacks trust by large sections of the public”. The Center for Countering Digital Hate analyzed several hundred X posts in the run-up to the 2024 presidential election and found that accurate community notes were not being displayed in three-quarters of cases. Misleading posts included claims that Democrats were bringing in illegal voters and that the 2020 presidential election was stolen. These posts received more than 2 billion views. A former UK minister says that use of AI to write notes can lead to “the industrial manipulation of what people see and decide to trust”. Elsewhere, Meta announced in January that it was getting rid of human fact-checkers in the US. Google announced fact-checks created by a professional organization would no longer be prioritized in search results because these were “no longer providing significant additional value for users”.