Tarento Group

Anthropic controlling metering has a lot of developers unhappy. I can't say I'm all that surprised, myself. Nor do I, even as an active AI developer, think it's a bad thing. Several things are happening at once, and if it happened with Anthropic, it will happen through the AI ecosystem. 🏭 We've maxxed out infrastructure. There aren't enough GPUs, enough chips, enough data centres, enough electricity. Rare earths are now caught in the middle of an unnecessary war, and that's having downstream effects globally. Experts have been warning about this for years, yet people are surprised when it happened. 📈 Almost nothing follows an exponential curve. Almost everything follows a logistics curve. Many performance improvements in AI lately have not been caused by improvements in LLMs directly, but rather by greater reliance at all levels on symbolic AI and algorithmic matching, with the AI serving primarily for orchestration, or by more effective use of data systems. These are generally mature costs, which means that token costs likely are stabilising in price. 💲 People who max out tokens are discovering this is extremely expensive. Using AI primarily to develop non-AI-dependent solutions (or even just foundations) then maintaining them manually is far more efficient. 🎥 This is happening throughout the field - AI image and video generation is now increasing in cost in most places, with rates fluctuating on a weekly basis. Model quality is still improving there (especially with video), but is also plateauing. 🫖 Most companies that bet the bank on connectionist AI downstream are now facing inverted economics which will likely bankrupt them. The ones that are succeeding for the most part are the ones that spent the most modernizing their deterministic data infrastructure; they've learned to sip rather than guzzle AI tokens, and use knowledge graphs and context graphs (holons) for the heavy lifting. ⌚ A context graph is just a fancy term for an event graph. A world model is a knowledge graph (what things are) coupled with a context graph (how things change over time). World models should inform AI, but not be stored in AI, because such storage is both lossy and too slow to update. An LLM is not a database. 🤖 AI is not going away, but nor is it in for a period of breathless exponential growth. Most of the money invested in it will not bring a return because physics matters. It will certainly change tech and society, but whether such changes will prove profitable has yet to be determined - this is the case with most technological innovation, fwiw.

There is an interesting recurrent pattern in technology. The people who invent, discover, etc., are very seldom the ones who benefit financially from their work. The reasons for it are simple, albeit not always intuitive: * Researchers in general are working in fairly rarified niche corners of a domain (or more often, at the intersection of domains) with work that even they will acknowledge doesn't have immediate payoffs. * The technological framework in which they are researching are typically still comparatively limited: most of the foundations of machine learning and neural networks, for instance, were first laid out in the 1960s and 1970s, but the technology was nowhere near powerful enough to do anything with it. * It takes time for a given idea to percolate through different layers of "experts", often with their own competing agendas or theories. The theory of plate tectonics bucked the orthodoxy of the day, and it would take decades for it to become mainstream because of this. * Until you build it and test it, you can't see all of the implications of a given discovery. Often the implementations are done by others, and they in turn also add to what is known about the particular invention or discovery. * Finally, there are the well-known social factors that may be involved - the person is the wrong gender, wrong race, wrong socioeconomic class, speak the wrong language, believe the wrong religion, etc. Bigotry is the single biggest impediment to innovation.

To the attention of my network.

Spent this morning at Glovis with a group of supervisors from varying departments. This group took the DISC assessment to measure their work styles and learn more about their natural leadership capabilities on a deeper level! This training gives them a better understanding of how to communicate, motivate, and resolve conflict within their own teams💪🏽🙌🏽 #leadershiptraining #DISC #alabamatrainingnetwork

I, like many others, have made the argument that LLMs need knowledge graphs in order to be grounded, but there are some caveats to that. LLMs were trained on, and work best with natural language structures - think sentence and story diagrams with sentences that have prepositional and adverbial clauses, have referential terms such as "this" and "that" . and have subordinate clauses that expand the core terms. These forms are all reifications - qualifications not just of single instances, but of relationships that connect one instance to another. Knowledge graphs have some implicit reifications, but for the most part knowledge graphs passed to LLMs tend to be fairly flat. By adjusting for reification, however, you can pass in graphs that compete well with natural language prompts. Explorations of language, LLMs, and knowledge graphs in this edition of … The Inferential Engineer!

Ah, the end may be near. Senior people leaving OpenAI is hardly new - remember the Thanksgiving 2024 coup when most of what would become Anthropic left? However, at the time, the company was still riding high in the media, and was still attracting investment. Now ... not so much. The big payday, the IPO, was when the big money would be made, and Sam could head off on his megayacht to compete with Elon on who could be the bigger TechBro. However, one misstep after the next has plagued OpenAI for months, and the possibility that OpenAI might go public has turned into a concern that OpenAI might not EVER get there, and that most of the investment will disappear into a puff of smoke (mainly $5K nVidia GPUs giving up the ghost in the few datacenters that managed to get completed).

We're beginning to see compounding effects in the AI field - symbolic AI is coming back in a big way as the need for grounding becomes important; the context window is becoming more optimised, the use of hybrid data storage and MCP/skills is pushing a lot of the memory into where it belongs - files and persistent stores. We have largely stopped chasing AGI (which I've ALWAYS felt was a waste of time) and started focusing on how to improve the technology in concert with, rather than replacing, sixty+ years of deterministic and algorithmic knowledge. We're also shifting the way that we think about applications; intentional program does not mean no coding; it means that you constrain what you want the application to do in as precise a manner as possible, refine it as you see that code in action, continue iterating in this way until you've achieved your objective, then you instantiate the code as an artifact (one that can run with only minimal intervention from an LLM) before looking for ways to improve it. It's a form of shared learning, but it is also something that is hard on the established software industry model.

I've worked with Enterprise Knowledge, LLC periodically over the years, even during the lean times when it seemed we were yelling out into the wilderness ( Joseph Hilger was a former boss of mine at Avalon Consulting), and I've talked with Zach Wahl and Lulit T. many times over the years at conferences and in conjunction with clients. There are a few key companies that have been doing the hard work of keeping the domain alive (as well as many in the semantic standards groups), including Semantic Arts, Inc. with Dave McComb , Michael Uschold and many others, TopQuadrant , Graphwise , Stardog , Neo4j and others. I bring this up in part because these are all people who have been solving many hard problems over the years in a sometimes esoteric domain, and as knowledge graphs and symbolic reasoning systems begin to get the exposure they deserve, these are also the people who have made a lot of this possible. We do not always agree - when you get a lot of very intelligent people in the same room, you also get highly opinionated people - but I think we're all involved in this for the same reason: this is necessary.

Here's a simple animated video showing how SHACL works as a two stage process, first selecting nodes, then for each node, evaluating the node against a property shape to determine whether it conforms for that shape. What is remarkable here is not the concept (though that's kind of cool), but the fact that the video was generated through the new Design mode in Anthropic Claude. I laid out a simple sketch showing the final frame, then created a prompt that walked through each step. Claude then took this description and created an HTML output (it could have been exported to powerpoint or Adobe PDF). I then used a screen capture to capture the animation as it ran and saved it as an MP4 file. Title time - about 10 minutes, and part of that was figuring out the bare bones interface. I did add the background music in a video editor (more for convenience, because I also needed to trim the video screen capture a bit), but the generation itself was very straightforward. Why is this important? I can generate static imagery fairly simply, but doing animations like this using powerpoint can be frustrating complex and time consuming, and even normal AI video creation from an image tends to be more hit and miss. This bridges that gap, and as a lot of the kind of work that I do would definitely benefit from animation, it opens up whole avenues for everything from training videos to cartoons.