Matthew Berman on Llama 2, AI Tools, Robots, and the LK-99 Claim

This week’s AI news ranges from Meta’s model releases and robots to new tools and a worrying cybercrime product. I also want to discuss an extraordinary claim about superconductivity, although it was still awaiting verification when this video was made in August 2023.

Meta was reportedly preparing AI chatbots for its social platforms, with different personas such as a surfer offering travel advice or a bot speaking like Abraham Lincoln. The reported goal was to increase engagement across its products.

Several uncensored Llama 2 variants had also appeared. In examples from Ollama, the standard model refused harmless prompts about a movie, spicy mayonnaise, and a hypothetical boxing match, while the modified version answered. I find these refusals frustrating: safety behavior can interfere with ordinary questions, although a handful of examples cannot establish that one model is generally better.

There is also a dispute over calling Llama 2 open source. The Open Source Initiative argued that Meta’s license imposes restrictions inconsistent with the open-source definition. In particular, a separate license is required when the licensee and its affiliates had more than 700 million monthly active users in the preceding month at the time of release.

Meta released AudioCraft for generating audio from text prompts. It presents possible uses such as helping musicians explore compositions and giving independent game developers an affordable way to create sound effects and ambient audio.

Meta was also reported to be working on labels for AI-generated content and message summaries. Its Open Catalyst demo applies AI to simulations of catalyst reactivity, with a claimed thousandfold speedup over existing computational approaches for that research task—not a blanket claim about every materials-science calculation.

YouTube was experimenting with AI-generated video summaries to help viewers decide what to watch. These would not replace creators’ descriptions. I wonder where the balance lies between giving viewers useful information and giving them so much that they no longer watch.

Stack Overflow, meanwhile, was introducing OverflowAI to use its programming knowledge base to provide answers. I have found myself asking ChatGPT about coding errors instead of going first to Stack Overflow, so I am interested to see how a service that has helped developers for years adapts.

A report described FraudGPT as a subscription service marketed for offensive purposes, including spear-phishing messages and other cybercrime-related work. The reported price was $200 a month. I am struck that the subscription software business model has reached even this corner of the internet.

AI-generated influencers were also gaining followers and brand deals. I worry that idealized images with no physical limits could intensify the pressures already created by manipulated social media pictures. That is my concern, not a measured mental-health finding; labels for generated content could help people recognize what they are seeing.

Models can increasingly use tools through API calls. I recently reviewed Gorilla, which aims to help a model use APIs and adapt when documentation changes. ChatGPT already had plugins, but the interesting direction is helping models work with tools they were not individually programmed to use.

I can imagine a general model orchestrating specialized models and asking each for help on its area of expertise. In my view this could be part of a route toward AGI, though that is a prediction, not evidence that these tools have already achieved it.

My AI-generated video of the week shows what looks like a robot competing in Power Slap. It is not real: someone used Wonder Dynamics to replace a human participant with a robot in existing footage. The result is convincing enough to illustrate how synthetic video can alter a scene.

Google’s RT-2 research brings vision and language models into robotics, translating learned concepts into actions in the physical world. Berman highlights its ability to generalize across tasks and use more complex reasoning prompts. These are research demonstrations of particular robotic capabilities, not proof of general intelligence.

The final story is LK-99. In 2023, researchers claimed that the material could superconduct at room temperature and ambient pressure. At the time I found the possibility exciting, while the claim remained under review and required independent verification.

I imagined possible consequences such as longer-lasting batteries, better electric vehicles and cooler chips. A superconductor can carry direct current without electrical resistance under suitable conditions, but that alone would not remove all heat or losses from a complete device, nor guarantee the specific improvements I speculated about.

Later evidence changed the picture: a November 2023 replication study reported that the electrical and magnetic anomalies attributed to LK-99 could arise from a structural transition in copper sulfide impurities, not superconductivity. The original room-temperature superconductivity claim was not established. See the researchers’ [study](https://arxiv.org/abs/2311.03558).