Lugh

The dream that LLM AIs might lead to AGI seems finally to be over. The idea's supporters said that with enough training and computation, independent reasoning would arise in LLMs. It never has. This week, OpenAI, the company that seeks a $500 billion stock market valuation, finally burst the bubble. It's long-awaited Chat-GPT 5, touted as 'near AGI', still makes all the basic errors in math and logic that an average 10-year-old wouldn't. Even Chain-of-Thought Reasoning, hoped by some to be an intermediary step, is a failure, too.

But LLM AIs aren't useless - in fact, the opposite. They've delivered Level 4 self-driving, are advancing general-purpose robotics, and excel at a wide variety of brain-work, even without AGI. That is enough to make them economically transformative.

But the US stock market seems bored with that. Instead, investors wanted to chase the possibility of a 'unicorn' - getting in early on a company that would achieve AGI, dominate like the other Big Tech giants, and earn trillions. That isn't going to happen any time soon.

China's economic system seems better at being pragmatic about AI. Most of their efforts seem geared towards useful real-world uses for AI, and much less about AGI. AGI will no doubt happen one day - who knows when? It will probably need fundamental breakthroughs developed by humans. Meanwhile, the real 2020s AI revolution is integrating AI into everyday life, not AGI.

A familiar 2020s story seems to be playing out here. The Chinese system of state capitalism seems to have grasped what matters when it comes to meaningful economic success, while the western system, driven by the super-rich becoming ever richer, seems bloated, inefficient, and unable to allocate capital & investment to what matters for society.

Here's a link to the paper - PDF 22 Pages

H5N1 Bird Flu hasn't gone away; in fact, the opposite. It's constantly spreading and becoming endemic in more and more animal populations. In North America, notably in the national dairy herd. All this increases the chances that a day comes when a mutation gives us a variant with 2 deadly characteristics. 1. Easily transmissible in humans & 2. A high mortality rate in humans.

mRNA technology is a bright spot in preventing future horror movie scenarios. It gives us the means to quickly develop a vaccine if a highly infectious and deadly variant arises. Amazingly, the US has just decided to dump that lifeline, and is jettisoning all funding for mRNA technology.

mRNA technology will continue to be developed in the rest of the world. Like more and more science and technology areas, China will probably become the leader. If the horror movie day comes, and a highly infectious and deadly human variant of H5N1 arises, Americans better hope their leaders are good at begging and pleading for help from the rest of the world in desperate circumstances, because they're going to need it.

US halt $500m in mRNA vaccine research, RFK says

I sometimes wonder if the future implications of neuromorphic computing are under-reported and discussed.

Neuromorphic chips have the potential to be true human/computer interfaces, in the way silicon chips just can't be. AI trained on neuromorphic computers may be more human-like, and very different from AI trained on silicon chips. If AI was to be integrated with a human brain, they would seem the most obvious candidate.

Finally, there's their fuel efficiency. That seems really futuristic compared to today's talk, from some AI leaders, of coal-fired AI data centers the size of Manhattan.

The world's largest neurocomputer simulates a monkey's brain

"We study subliminal learning, a surprising phenomenon where language models transmit behavioral traits via semantically unrelated data. In our main experiments, a "teacher" model with some trait T (such as liking owls or being misaligned) generates a dataset consisting solely of number sequences. Remarkably, a "student" model trained on this dataset learns T. This occurs even when the data is filtered to remove references to T."

This effect is only observed when an AI model trains one that is nearly identical, so it doesn't work across unrelated models. However, that is enough of a problem. The current stage of AI development is for AI Agents - billions of copies of an original, all trained to be slightly different with specialized skills.

Some people might worry most about the AI going rogue, but I worry far more about people. Say you're the kind of person who might want to end democracy, and institute a fascist state with you at the top of the pile - now you have a new tool to help you. Bonus points if you managed to stop any regulation or oversight that prevents you from carrying out such plans. Remind you of anywhere?

Original Research Paper - Subliminal Learning: Language models transmit behavioral traits via hidden signals in data

Commentary Article - We Just Discovered a Trojan Horse in AI

It's still early days, and the test was only on 53 people, but a new drug called Trontinemab almost completely eliminated the brain plaques indicative of Alzheimer's in 91% of them. Wider trials on 1,800 people will take place later this year. Fingers crossed. Alzheimer's is dreaded by many people; a cure or near-cure would have a major impact.

Roche’s New Alzheimer’s Drug Trontinemab Nearly Eliminates Brain Plaques

Form Energy in the US is also developing this technology, though they haven't deployed to the grid yet.

As electricity grids get nearer to being 100% renewables, they need to account for <5% of times both solar & wind don't meet peak electricity demand. Lithium-Ion batteries, which only store electricity for a few hours, aren't much use here, but Iron-Air batteries will be.

They can store days worth of electricity, and not only that, they are stable and non-flammable. The only chemical reaction taking place is iron oxidizing (rusting).

Ore Energy connects world’s first grid-connected iron-air battery in Delft

"We're not planting our flag and leaving. We’re going to stay, learn, and then go to Mars. There’s critical real estate on the Moon. We want to claim that real estate for ourselves and our partners, which is going to be critical to being successful in that mission."

Sean Duffy interviewed this morning on NASA+.

The Outer Space Treaty, which 117 countries, including the US, are signatories to, prohibits Earth nations from claiming lunar territory. The trouble with saying you can break any international law you want, by say, invading Greenland, or claiming the Moon, is that then anyone else can. By say, invading Taiwan, or claiming the Moon, also.

What do you do then, especially when they (China) get all the good bits of the lunar South pole first? Chinese plans for their International Lunar Research Station are far more advanced than anything NASA has. There's every likelihood they'll be the ones able to claim best the lunar real estate first.

Interesting article McKinsey’s thesis is that foundation models (think vision‑language‑action brainpower) let robots recognize objects, follow spoken commands, and behave flexibly. Imitation learning and behavioral cloning let them watch humans and learn movements without explicit programming.