Martin Vechev

Switzerland just dropped a bombshell: ETH Zurich and EPFL are launching a fully open-source Large Language Model that will expose its entire training data, source code, and decision-making processes to public scrutiny. Unlike ChatGPT or Claude, where the inner workings remain corporate secrets, this Swiss model promises radical transparency. Imagine being able to trace exactly why an AI made a specific decision, or having complete visibility into potential biases in training data. The model will support over 1,000 languages and run on Switzerland's Alps supercomputer, positioning it as a serious alternative to Silicon Valley's black boxes. https://lnkd.in/eYCVKDzc #AI #tech #law #wolftheiss

129

9 Kommentare

I spent a weekend testing different AI agents (ChatGPT and others) as physics assistants. First, I generated a truly original idea, one that I am confident is not yet known, and confirmed that neither Google nor the AI agents had any knowledge of it. Then I asked the AI agents to perform step-by-step calculations with detailed instructions, corrections, and follow-ups. At every stage, the AI made mistakes in calculus, linear algebra, and differential geometry, but with guidance it was able to correct them. In total, I accomplished in a single weekend what would previously have taken at least a month, if not longer. Here are my conclusions and predictions in the long run: 1. The future of fundamental research in physics lies with AI agents. We are on the verge of having tools that are genuinely useful for fundamental physics, and they will soon be developed. Either you learn how to use these tools, or you are out. 2. The amount of garbage generated by AI will grow exponentially. The hallucination problem is fundamentally impossible to eliminate, and the resulting mistakes could be difficult to detect. Either you learn how to identify and correct mistakes, or you are out. 3. AI agents based on transformers will never generate new ideas. They can enable us to borrow methods across disciplines to solve different problems, but that is all. Either you learn how to generate ideas yourself, or you are out. 4. Large collaborations of experts will no longer be necessary. If knowledge outside your immediate area of expertise is required, the most effective solution will be to become an expert yourself. Either you learn how to learn quickly, or you are out. 5. Smaller collaborations will become essential. The central challenge will be generating new ideas, and for that you actually need to brainstorm with colleagues. Either you learn how to clearly and quickly communicate your ideas, or you are out. Unfortunately, I do not think the physics community is ready for this change, and my prediction in the short run is that it will divide into three sub-communities: 1. Traditional physics. This group has the resources and skills to continue generating ideas and carrying out calculations the traditional way. 2. Vibe physics. This group will begin doing “vibe research,” relying heavily on results generated by AI, often without thorough verification. 3. Hybrid physics. This group will embrace AI as a powerful tool while maintaining high standards for generating ideas and performing calculations.

105

54 Kommentare

Oh, Dear! Yet more magical thinking about reducing hallucinations of LLMs by using diffusion models. As Denis O. critiqued them, this won't happen. Diffusion models are at least as much subject to hallucinations as transformer based LLMs. Anyone who has used a GenAI image generator knows of the difficulty in getting it to faithfully follow the prompts. "It seems likely that diffusion LLMs could challenge auto-regressive tech, offering new capabilities like improved reasoning and controllability, but their full impact is still emerging." As always, the proponents of these new technologies keep talking about reasoning. It is just not going to happen and we know that image generators are highly uncontrollable. The idea of generating a whole block of text in parallel is wrong from a semantics and linguistics perspective. Language has sequential dependencies. An argument is developed sequentially. A critical evaluation is developed sequentially, not in parallel. "Andrej Karpathy and Andrew Ng, both renowned AI researchers, have enthusiastically welcomed the arrival of Inception Lab’s diffusion LLM." Surely, this is the kiss of death, not an endorsement. "This process begins with random noise which is then gradually refined and “denoised” into a coherent stream of tokens. This is analogous to how diffusion models generate images by starting with noise and iteratively removing it to reveal a clear image." It is not analogous, it is exactly the same, need to get the facts right. "Parallel Processing: Diffusion models can process and generate text in parallel, potentially leading to significant speed advantages." As with image generators, the final output of tokens will be fast, just transferring out of the output buffer to the user, but after a long pause (not even thinking) to do all the denoising. We also know that diffusion models eat TFLOPs, just as much as or more than transformers. https://lnkd.in/eq2HWhWZ

73

25 Kommentare

Ed Zitron, crystal clear as always: "[T]he real problem is that, just like big tech’s new gluttony of physical real estate it's taken on, generative AI companies are burdened with a constant and aggressive form of cloud debt — the endless punishment of the costs of accessing the API for generative AI models that always seem to get a little better, but never in such a way that anything really changes other than how much Anthropic and OpenAI are going to need at the end of the month or they break your startup’s legs. [...] The literal only liquidity mechanism (outside of Cognigy) that generative AI has had so far is “selling AI talent to big tech at a premium.” Nobody has gone or is going public, and if they are not going public, the only route for these companies is to either become profitable — which they haven’t — or sell to somebody, which they do not. But I’ve been dancing around the real reason they won’t sell: because, fundamentally, generative AI does not let companies build something new. Anyone that builds a generative AI product is ultimately just prompting the model, albeit in increasingly more-complex ways at the scale of something like Claude Code — though Anthropic has the advantage of being one of the main veins of infrastructure. This means that a generative AI company owns very few unique things beyond their talent, and will forever be at the mercy of any and all decisions that their model provider makes, such as increasing prices or creating competing products." Let me repeat this for those in the back, in case it was not clear enough: "This means that a generative AI company owns very few unique things beyond their talent, and will forever be at the mercy of any and all decisions that their model provider makes, such as increasing prices or creating competing products." In other words, all that fanfare about innovation and progress and co. is born beheaded when it must depend on those who own the infrastructure, the data, and the products others depend on. Why are some educational institutions and educators so naive? Why they don't want to see that the intellectual and financial tithes they are paying will only go up? https://lnkd.in/e7vvFsE6

79

2 Kommentare

Parallel agents are emerging as an important new direction for scaling up AI. AI capabilities have scaled with more training data, training-time compute, and test-time compute. Having multiple agents run in parallel is growing as a technique to further scale and improve performance. We know from work at Baidu by my former team, and later OpenAI, that AI models’ performance scales predictably with the amount of data and training computation. Performance rises further with test-time compute such as in agentic workflows and in reasoning models that think, reflect, and iterate on an answer. But these methods take longer to produce output. Agents working in parallel offer another path to improve results, without making users wait. Reasoning models generate tokens sequentially and can take a long time to run. Similarly, most agentic workflows are initially implemented in a sequential way. But as LLM prices per token continue to fall — thus making these techniques practical — and product teams want to deliver results to users faster, more and more agentic workflows are being parallelized. Some examples: - Many research agents now fetch multiple web pages and examine their texts in parallel to try to synthesize deeply thoughtful research reports more quickly. - Some agentic coding frameworks allow users to orchestrate many agents working simultaneously on different parts of a code base. Our short course on Claude Code shows how to do this using git worktrees. - A rapidly growing design pattern for agentic workflows is to have a compute-heavy agent work for minutes or longer to accomplish a task, while another agent monitors the first and gives brief updates to the user to keep them informed. From here, it’s a short hop to parallel agents that work in the background while the UI agent keeps users informed and perhaps also routes asynchronous user feedback to the other agents. It is difficult for a human manager to take a complex task (like building a complex software application) and break it down into smaller tasks for human engineers to work on in parallel; scaling to huge numbers of engineers is especially challenging. Similarly, it is also challenging to decompose tasks for parallel agents to carry out. But the falling cost of LLM inference makes it worthwhile to use a lot more tokens, and using them in parallel allows this to be done without significantly increasing the user’s waiting time. I am also encouraged by the growing body of research on parallel agents. For example, I enjoyed reading “CodeMonkeys: Scaling Test-Time Compute for Software Engineering” by Ryan Ehrlich and others, which shows how parallel code generation helps you to explore the solution space. The mixture-of-agents architecture by Junlin Wang is a surprisingly simple way to organize parallel agents. [Truncated for length. Full text, with links: https://lnkd.in/gQ98HMci ]

2.473

174 Kommentare

🔮 Why AI4Science is hard: open the physics challenge? In the past few months at conferences and seminars at Stanford, Cambridge, and Yale I’ve been asked a lot: what are the real use cases for ML in experimental physical sciences - and why is it so difficult? Indeed, the headline successes over last several years - symbolic regression “rediscovering” Kepler laws, SINDY learning Newtonian mechanics or learning ODEs/PDEs on clean simulation and real world turbulence are inspiring, but transition to the physics labs seems to be elusive. In fact, my own journey into electron microscopy ~15 years ago started from realization that scanning probe microscopy data is very difficult to analyze by ML. At that time, we interpeted these difficulties as "calibration" and "uncontrolled factors". Now the word will be distribution shifts and hidden latent variables, but the central premise is the same. Conversely, in electron microscopy atom is an atom - what you see is what you analyze. Hence, we can think about “learning physics” from microscopy as e.g. fitting force fields from snapshots of atomic configurations. However, the reality is still more complicated - between observations and underpinnign physics there is a thick layer of physics abstractions. In scanning probe and electron microscopies, most phenomena are strongly entangled (ferroelectricity and electrochemistry for PFM, electronic and structural order for STM, and so on), and the physics is only partially known. After interetsing discussions with Georgia Channing, I have meditated on what would a good discovery benchmarks look like? Here’s a concrete one: STM on α-RuCl₃, a honeycomb magnet with Kitaev-type (bond-dependent) interactions. Briefly, Kitaev materials are magnets where the atoms sit on a honeycomb-like lattice and their spins interact in a very unusual, direction-dependent way: the type of spin coupling depends on which bond connects two atoms. This behavior comes from strong spin–orbit effects in certain crystals made of heavy elements. In the ideal case, these materials can host a “quantum spin liquid,” a state with unusual magnetic order and exotic fractional excitations. Now, this is the idealized physics. The task for AI: how much of this physics can be discovered or hypothesized from an STM image showing unusual non-periodic patterns of electronic density? How can prior physics context or simulation help understand these phenomena? Now, over my ~30 year career in scanning probe microscopy, I can probably identify 5-6 examples like that from my and other's work (chaos in ferroelectric switching, Berry phases from STM, few others) and several examples where the answer is still out. But may be this is a start! If we want ML that discovers rather than denoises, we need community benchmarks like this - where success means suggesting mechanisms and proposing the numerical model. You can read more about RuCl3 here: https://lnkd.in/eeRbFVDG

101

12 Kommentare