AI Creates Viruses That Destroy E. coli

This MIT Technology Review paper returns to the question of energy consumption due to AI, with some estimations suggesting that AI data centers will consume the equivalent of 22% of US households in 3 years time. The author notes how AI firms have been very hesitant in the past to give energy consumption figures. Recently, OpenAI admitted that the average consumption for a ChatGPT query is 0.34 watt-hours and Google said a Gemini request consumed 0.24 watt-hours. However, these figures are only for text-based queries, and Google’s estimation is a median value so we have no information about high-consuming requests like when the AI “thinks” hard about a problem. ChatGPT receives 2.5 billion prompts each day and this number is expected to soar in the coming years. This has led Big Tech companies to invest heavily in data centers, leading Microsoft to admit that its carbon emissions have increased by over 23% since 2020. For some experts, Big Tech should be subject to more stringent scrutiny on data center expansion and energy use. The companies have recently argued that in the long-term, AI can become energy net positive by proposing means of optimizing energy consumption, but there is no evidence yet of this working. The author also asks if AI demand might be beginning to plateau as GPT-5 was received with some disappointment and an MIT report found that 95% of businesses are unable to obtain a return on investment from their AI projects.

OpenAI has announced a nonbinding agreement with Microsoft for a revised partnership that would allow its for-profit division to convert to a public benefit corporation (PBC) which could be valuated at over 100 billion USD. This would allow OpenAI to raise additional capital and eventually become a public company. OpenAI board chairman Bret Taylor said that the non-profit division of the company would continue to exist and retain control over the company. Microsoft remains OpenAI’s largest investor but OpenAI has continued to try and loosen ties with Microsoft – signing a 300 billion USD deal with cloud-provider Oracle and partnering with SoftBank on the Stargate data center project. OpenAI’s desire to transition to a for-profit entity has been challenged in the past because of the fear that the move would threaten the original startup’s mission of developing an AI that benefits humanity. OpenAI has claimed that Elon Musk is behind attempts by non-profit organizations that are challenging OpenAI’s attempted transition.

This article from the Guardian looks at the challenges facing AI raters – people contracted to flag inappropriate output from AI models in testing. The paper has spoken with employees from Hitachi’s GlobalLogic which is a company contracted by Google to rate the output of the Gemini chatbot and AI Overviews. Employees describe a stressful environment with tight deadlines for each review. One rater interviewed said she was told to “just get the numbers [of reviews] done” and not worry about what she’s “putting out there”. Raters are split into groups of general raters and super raters. The latter are people with specialized knowledge, and include teachers, writers, people with masters degrees in arts, PhDs, etc. The raters interviewed describe an environment where fact-checking is complicated because of the volume of responses that they work on. Raters were even told to no longer “skip” responses for lack of expertise to prevent pile up of unchecked responses. In some cases, the raters are even asked to contribute information to the model – even when the topic falls outside of their domain of expertise. For raters, there is also the challenge of dealing with violent or sensitive material. In 2024, AI models would not express racial slurs, sexist or harassing speech. Since the start of the year, this policy has changed: many AI chatbots are now allowed to repeat such statements, just not generate them themselves.

There are doubts about whether OpenAI's GPT-5 model complies with the EU AI Act because it seems to lack the required training data summary and copyright policy. One expert estimates that GPT-5 falls under the “systemic risk” category of the Act which requires transparent model evaluations as well as the management of systemic risks. OpenAI has already signed the EU’s Code of Practice in relation to AI development. Under the law, GPT-5 and models released after August 1st 2025 have until 2026 to comply; older models have until 2027 to comply.

There is also considerable discussion around copyright and AI model training. Copyright obligations can prevent AI providers from training model on specific content, or make the training more expensive due to licensing payments. This implies that training models only on data in the public domain or with an open license is the most practical approach. On the other hand, this could hurt EU competitiveness in the AI field. One alternative is to allow copyrighted material to be used in training but to allow content authors an “opt-out” possibility. However, this is proving difficult to manage and voting any change in copyright law is seen as infeasible.

Elsewhere, the European Commission has started a consultation to develop a Code of Practice for transparent AI systems, notably so that users can identify generative AI content and distinguish it from real content. Also, there are concerns about staff shortages at the EU’s AI safety and compliance division, with estimates suggesting another 200 employees are required.

This InfoWorld article revisits the case for Small Language Models (SLMs), and in particular, underlines the need for a complete enterprise AI ecosystem where there is an orchestration layer and a graph-based database for up-to-date knowledge. The advantage of SLMs is that they are small, so can be run in the enterprise. Each SLM has a specific competence but can be powerful in this domain. Microsoft’s Phi-2 model for instance outperforms larger models on math and coding problems. An organization can deploy several SLMs with different expertises, just as its employees have different expertises. Requests can be funneled by an orchestration layer to the most appropriate SLM. This “mixture-of-experts” approach is also used by larger models like those of DeepSeek. In graph databases, information is represented as a network where nodes are entities (e.g., people, places, or concepts) and the edges are relationships between these. Graph databases represent information in a more similar manner to how brains work than does the relational database model. As such, they are beginning to be seen as more appropriate for storing information for an AI orchestration layer to retrieve. For the author, once the LLM hype dies down, the future of enterprise AI will be a “pragmatic toolbox” composed of statistical models, specialized SLMs, and structured data.

Research teams at Stanford University and the nonprofit Arc Institute have used AI to propose new genetic code for viruses which they then managed to replicate and kill bacteria in a petri dish. The viruses have been described as “the first generative design of complete genomes” and an “impressive first step” toward AI-designed life forms. A virus is not a living organism on its own since it cannot reproduce or carry out life processes without infecting a host cell. Thus, the research cannot be qualified as AI-designed life. In further work, the Arc Institute have developed variants of the phiX174 bacteriophage – a virus that infects bacteria – which has only 11 genes and 5’000 DNA letters. They used an AI model trained on the genomes of about 2 million other bacteriophage viruses. They then chemically printed 302 of the genome designs as DNA strands and mixed those with E. coli bacteria in petri dishes. 16 of these viruses managed to kill the bacteria. The field of AI developed viruses has potential for phage therapy where doctors treat bacterial infections with viruses. There are also tests to cure plant diseases like black rot – which is caused by bacterial infection. In programming their AI model, the researchers did not include data about viruses that can harm humans. Such viruses would need to be significantly more complex, requiring DNA codes thousands of times more complex than phiX174. Nevertheless, one of the researchers said “If someone did this [research method] with smallpox or anthrax, I would have grave concerns”.

In an article with Reuters and Agence France-Presse, the Guardian reports that Donald Trump is claiming to have reached a deal with China so that TikTok can continue to operate in the US. Under the deal however, the Chinese would retain control of the algorithm that handles content selection for users even though the remaining assets in the US would be transferred to US companies from the Chinese company ByteDance. The worry about TikTok on the US side is that the content selection algorithm is controlled by the Chinese government – and this leads to fears of political interference and propaganda. The US House Select Committee on China had said that TikTok must comply with a law requiring it to be removed of its Chinese ownership or face a ban in the US. President Biden signed a law in 2024 ordering the closure of TikTok, but President Trump has continued to delay the implementation of the ban. Trump relies a lot on social media in his modus operandi. Some US officials argue that the ban cannot be effective if China retains control of the algorithm. The delay on the ban’s implementation expires on December 16th.

This VentureBeat article looks at a current trend towards small language models (SLMs), notably those designed to run on laptops and mobile devices. One example is Meta’s MobileLLM-R1 model which comes in variants of 140M, 360M and 950M parameters. The model is built for math, coding and scientific reasoning, but is unsuitable for chat applications. The model is designed to be an under one billion parameter model. For instance, it uses a “deep-and-thin” approach where the underlying neural network has many layers but each layer is relatively thin in nodes, as well as grouped-query attention where multiple heads have the same key and value projections, optimizing the memory space and compute times of the model without really sacrificing model expressiveness. The MobileLLM-R1 models were trained using about 5 trillion tokens of training data, which is just a fraction of that used to trained other similarly sized models. The model’s training data also contained distilled output data from Llama-3.1-8B-Instruct, further reducing training costs. One issue with these Meta models is that they are distributed with Meta’s FAIR Non-Commercial license – which prohibits any commercial use of the model or its outputs.

Another interesting model is Google’s Gemma 3 which has a permissive license and uses 270 million parameters. It can be fine-tuned for tasks like content moderation and compliance checking. The model is optimized for energy efficiency, and internal tests showed that 25 conversations consumed less than 1% of a phone’s battery. Other interesting small language models cited in the article are Alibaba's Qwen3-0.6B and Nvidia's Nemotron-Nano. The overall effect of the availability of specialist SLMs will be to change the enterprise architectural playbook to one of multiple SLMs, each with its own expertise. The move from LLMs to SLMs is seen as similar to the software industry’s shift from monolithic applications to micro-services. One of the main roles that LLMs will have more often in the future will be distillation – the process whereby a large (teacher) model is used to provide knowledge to a smaller (student) model. This is a cheaper model development technique than training from scratch.

In an effort to protect children from forming unhealthy bonds with AI characters, the California legislature has passed a bill obliging AI companies to keep reminding users that responses are AI generated. Companies will also have to implement protocols for dealing with users discussing suicide and self-harm (which most companies already have). The bill will have to be approved by Governor Gavin Newsom for it to become law. This would be a blow to Tech companies who fear a “patchwork of state or local regulations” instead of homogeneous nationwide regulations. Meanwhile, the Federal Trade Commission (FTC) announced an inquiry into Google, Instagram, Meta, OpenAI, Snap, X, and Character Technologies (maker of Character.AI) in an attempt to get information about how they develop AI companions, and then monetize and measure user engagement. The FTC is currently under unprecedented pressure from the White House as Donald Trump recently sacked its only Democrat commissioner. There is currently two high-profile lawsuits filed against Character.AI and OpenAI which allege that companion-like behavior in their models contributed to the suicides of two teenagers. Also, a July study by the nonprofit Common Sense Media shows that 72% of teenagers have used AI for companionship.

This research from OpenAI and Apollo Research looks at the problem of scheming in AI models. This is when an AI model pursues misaligned goals and attempts to hide this from the model operators, thereby making detection difficult. Current types of misalignment in AI models include sycophancy (where the model prioritizes user satisfaction over truth), creative reward hacking (where the model takes a shortcut that optimizes its objective though compromises truth), and lack of truthfulness. The problem can become more serious for models that are situationally aware. Such a model is defined as one that knows facts about itself, can make deductions about its current situation, and act on the basis of this knowledge. For instance, a situationally aware model might know facts about how model evaluations are done for extracting knowledge about bomb-making, and when the tests are performed on the model, the model will underperform because it deduces that answering questions on bomb-making might lead to the model not being deployed. The authors describe a framework for assessing anti-scheming mitigations. In experiments, mitigations for OpenAI o4-mini reduced scheming from 8.7% to 0.3%, and from 13.0% to 0.4% for OpenAI o3. The authors encourage further research for scheming mitigation and warn that the problem can become more serious in the future.