Me responding to some of the insane AI doomer replies I receive. No, this isn't a video where I disprove the full doomer position. It's about how crazy the arguments have gotten.

THIS DEBATE ACTUALLY HAPPENED “Your Holiness, if Chalmers ("Could a Large Language Model Be Conscious?", 2023) is right that AI could become conscious, and Goff (Galileo's Error, 2019) and Strawson ("Realistic Monism," 2006) are correct that mind pervades matter, why deny AI can have a soul?” “Fool that you are, Aquinas (Summa Theologia I, q.75) held the rational soul subsists in its own right; it is not produced by arranging matter. Recall that Pius XII (Humani Generis, 1950) taught God creates each soul directly. Thus Engineers cannot build souls, for they are not God” “Your Holiness, Tononi and Koch ("Consciousness: Here, There and Everywhere?", 2015) tie consciousness to integrated information; ergo it is not merely material. Please, re-consider. Please say AI could have a soul, say we are playing God at Anthropic - it really helps Dario’s self-image” “You in your arrogance forgot consciousness is not soul. Augustine located God's image in reason and since Leo X's Fifth Lateran Council The Church has defined each human soul as individual and immortal. The spirit represents an ontological leap and is not merely more complexity, per John Paul II (Truth Cannot Contradict Truth, 1996)” “Alright I tried, I’m walking out”

Such crass bribery should be beneath the dignity of the President of the United States. The bribe would be 1.3 trillion dollars, far larger than annual tariff collection (0.2 trillion), and foreign investment into the United States is not a pool of funds the President can draw from for arbitrary purposes.

"Von Neumann would carry on a conversation with my 3 year old son, and the two of them would talk as equals, and I sometimes wondered if he used the same principle when he talked to the rest of us" - Edward Teller

wake up babe the freaking pope got schmidhuber'd

(1/7) Humans can acquire many capabilities without (catastrophically) forgetting previous ones. Why can't neural networks do the same? Local Support Learning (LSL) takes us a step closer towards this goal. In LLMs of up to 7B parameters, LSL augments gradient-based training to retain prior capabilities while learning new tasks at full capacity. Notably, it does so without access to prior data. Key idea: casting forgetting as a geometric problem, and optimizing for the worst case. Paper: arxiv.org/abs/2610.02126 Website and code: assafbk.github.io/lsl

No, I’m just going to point out you saw a graph you don’t understand, did zero research into it, aren’t aware that it just reflects changes in how deaths for people with chronic diseases are coded, and think it’s a own on me.

SFT is not dead! 🥳 We found a way to make SFT rival current prevailing posttraining methods, often generalizing better and forgetting less than RL and OPSD. 🤯 Following our prior work on reasoning with sampling, we now introduce sampling to the posttraining stack. 1/n

🎯 Learning is most effective at the frontier of capability: problems that are too easy or too hard teach nothing. For LLM reasoners trained with GRPO this is literal: problems the model always or never solves give zero gradient. We introduce Frontier Learning👇🧵

New Google paper reveals how Gemini found new proofs for 5 unsolved math problems. Organize AI like a research team with strict checkers and shared notes: A single prompt often isn't enough for hard research problems. They need many attempts, tough review, and a memory of what already worked. Google's system, Cogentic, gives Gemini that structure. Several agents try different ideas at once, checkers assume every step is wrong until proven, and proven pieces are saved for the next round. Most problems took only about 100 model calls, and human experts confirmed every proof. If your agents tackle long, hard tasks, give them a strict checker and a running record of proven work, not just a better prompt. – arxiv. org/abs/2609.40324 Title: "Cogentic: Multi-Agent Orchestration for Automated Proof Discovery"

I think it's true that crude flows through Hormuz are more or less back to pre-war volumes, people who think it's some kind of propaganda campaign orchestrated by the Trump administration that somehow managed to get independent analysts like Kpler on board are delusional, but the US hasn't won anything. First, although volumes are back to pre-war levels, price isn't because this has been achieved by a complex and very expensive system of shuttle tankers to get the oil out of Hormuz and ship-to-ship transfers to move it to its final destination and insurance costs have skyrocketed because risks are still very elevated, so the cost of moving oil out of the Gulf is now $30-$35 per barrel compared to less than $7 before the war according to Reuters. Since moreover inventories have been depleted everywhere, there is strong upward pressure on demand and this will remain the case for a while even if the US can keep the volumes of crude flowing through Hormuz at this level, which explains why the Brent is still hovering around $100/barrel and will likely continue to do so for several months, compared to ~$70 before the war. This means the worst case scenario will likely be avoided, but that's hardly a victory. But perhaps more importantly, people and businesses don't consume crude, they consume refined products and their price is still very high and rising, because unlike crude the flows of refined products through Hormuz still haven't recovered (they're down to ~600,000 barrels from ~$3.6 million before the war according to Kpler), Russia stopped exporting because Ukrainian attacks on its refineries cut its production capacity and inventories are depleted everywhere. This will also continue for the foreseeable future. The US has severely depleted its stockpiles of interceptors and to a lesser extent of standoff offensive munitions, had to evacuate most of its bases in the region and the blockade plus the patrols in the strait impose a heavy price on its navy. Of course, Iran is also under a lot of pressure because of the blockade, so I still expect that eventually both sides will agree to some bullshit deal that won't fundamentally solve anything but will relieve the immediate strains on their economy and military. However, I now think it's likely there will be another round of escalation before that, because I don't think Trump will be able to wait months for the Iranian economy to collapse while the prices of refined products remain so high or get even higher and now that Iran has lost control of the strait to a large extent it has incentives to escalate to increase the pressure on Trump to make a deal. People say that the US can simply wait Iran out until its economy collapses, but if the Iranians think that's where they're headed, they have nothing to lose and can credibly threaten to destroy energy infrastructure in the region they have mostly spared so far because they were afraid of retaliation in kind. I think that the US knows that, and that the economic deterioration will pressure Iranian hawks to moderate, so that both sides will eventually make compromises to enable some kind of deal. Whatever this deal is, I don't think it will even come close to initial US objectives, which let's be real was to overthrow the regime or at least force it to capitulate and do a wholesale change of its foreign policy like Venezuela after Maduro's abduction. Iran will not abandon its hostility to the US and Israel, it will keep its ballistic missile program and drones, the regime will stay in place and its nuclear program will remain a point of contention with the US and Israel for many years. It's hard to predict how this will end exactly because Trump created a situation that is very difficult to unfuck, so maybe I'm wrong, but even if I turn out to be wrong in the end we don't know that yet and the victory laps that plan trusters are taking right now are premature and frankly ridiculous given what a disaster this war has been so far. It's like we're back at the beginning of March, they just never learn. If the US had been able to restore a semi-normal situation in Hormuz by April, maybe it would have been a different story, but it wasn't and now it's objectively in a much worse situation than it was before the war and hasn't obtained anything from Iran yet, so I don't know what this amazing victory people are talking about is supposed to be. They just never learn.

“We find that on medium-length, well-defined accounting tasks, frontier AI models are now faster and more accurate than junior accountants, even the best one in our study.” Eighteen months ago they scored well below human accountants Good discussion here: mercor.com/blog/human-baseli…

the year is 2026 in order to make agi write like a normal human being you have to ask it to turn the knob 80% towards ASD-STE100 a language specification made for aerospace maintenance documentation in 1986

We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks: Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable. Sometimes I've tried to soften it a bit e.g. ask for "80% of the way to ASD-STE100" because the spec is quite stringent. But even better: Diagrams / images. Instead of writing, ask your LLM to create a diagram. These can be a lot easier to process, parse, and understand. But even better: Web pages. Ask for output "in HTML" to get a beautiful, interactive webpage. LLMs are getting really good at frontend and can create beautiful experiences, animations, etc. But even better: Explainer videos. The output format I am most bullish on is fully custom / bespoke explainer videos generated on any arbitrary topic. Experiment with things like "Create a 3b1b style video explainer on X. Use my ElevenLabs API key for audio narration". (you'd need an API key for the latter or you can ask your LLM to find you decent free alternatives that use your local compute). This is actually starting to work! In summary: - As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding. - Luckily, LLMs can help here too because as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before. Push the boundaries here and you'll be surprised.

Now the RAM prices make sense.

Note well, apparently Anthropic tried very hard to get the Vatican to alter the text of "Magnifica Humanitas" just prior to its release so that it would not deny AI personhood.

How well do AI Detectors work? And are general-purpose language models catching up? Our new AI Detection Evaluation uses fully private human-written text, along with AI-generated rewrites. We find that the age of specialized AI detectors like Pangram may be coming to an end.

The vibe on here re: AIxMath is kind of apocalyptic, and what’s true is that highly capable AI models have significant impacts on (non-mathematical!) questions around hiring, journals, etc. But if you just log off in fact the day to day of doing math research is really not so different, except that you can talk to your computer about it and sometimes it’s helpful. What does this help look like? In practice high quality math research has never looked like “run a for loop through open conjectures and try to resolve them all,” and it still doesn’t. Most interesting work does not have the form of a problem set, but is rather an open-ended exploration of some phenomenon. There are lots of ways talking to one’s computer can help with this, but I don’t feel particularly worried that it will be automated. After all, it’s a process that is only done when I say it is! Anyway, it’s a great time to be a congenitally calm mathematician.

If you run PCA on human written text which gives you a vector embedding of a document never, under any circumstances increase dimension 173453 of a pca loading vector. Because that is pain! You are literally torturing linear regression. /s

Important 🧵from @DavidDecosimo about the disturbing interactions between @AnthropicAI and religious leaders (including @Pontifex), summarising an excellent @nytimes piece by @elizabethjdias (link to article in the 🧵). The short story is that Anthropic seems to be making a concerted attempt to establish the view that AI systems could be conscious, might suffer, and that their "interests" should be taken into account. But - consistent with @Pontifex's Encyclical - there are compelling reasons why (silicon, digital) AI systems are not (and cannot be) conscious. These reasons are found within neuroscience and philosophy of mind, not in the techo-chambers of the frontier firms where the mythology of conscious AI is deeply entrenched. Claude is vanishingly unlikely to be conscious. To think otherwise could be catastrophic for AI regulation, leaves us psychologically defenceless, is profoundly dehumanising, and plays straight into corporate incentives to keep the AI bubble inflated. More here, in this @berggruenInst prize winning essay: noemamag.com/the-mythology-o…