Week 25

"@pikesley@mastodon.me.uk

Formulating an #EMFCamp policy: if I'm in your talk and your slides have #LLM slop on them, I'm walking out. No ifs no buts. You couldn't be bothered making them, I'm not interested in looking at them, and I have to presume that whatever you're saying has had the same amount of thought put into it"

LLM complemented me for catching the error 😂

When I confronted it, it said this:

"You're right, I was sloppy there... LLMs generate text token by token, and math is particularly dangerous because a plausible-looking intermediate step is easy to produce without actually checking it. I filled in a step confidently that I hadn't carefully verified..

Even when the source material is right there in the context.. I'm still generating tokens based on what seems likely to follow, so my own learned patterns can override what's literally in the provided text.

In this case the telescoping idea was in [textbook] derivation, but the specific intermediate algebraic steps weren't spelled out [wants to be spoonfed?] explicitly even there — so I had to generate them, and that's where my pattern-completion drifted from correct algebra.

It's a fundamental limitation to be aware of.. I'm not doing symbol manipulation [basically math the way humans do it], I'm doing sophisticated pattern matching that usually looks like symbol manipulation."

Oh boy, "frontier" LLM made a massive algebraic mistake..

Now I'm not even sure they can be used for ed assist.

There were some other reports too saying Palantir software isn't that good. But Focker somehow created an aura of greatness about the product. Focker is always in the circle of trust 👌.

"@ChrisMayLA6@mastodon.me.uk

You may recall I posted recently on the highly focussed character of the claimed productivity benefits of Palantir's technology in the NHS (mostly happening in one London hospital); now NHS England has quietly admitted that it actually does not have robust evidence to link any wider rises in efficacy & performance to Palantir's software....

So, just to be clear; their claims are just than, claims unsupported by evidence - which further suggests political interference!"

NYT: "The president of the Utah State Senate, who championed a huge data center beside the Great Salt Lake, was defeated in his Republican primary on Tuesday night, one of the most high-profile signs of the voter backlash to data center projects.

The vote to oust the Senate president, J. Stuart Adams, was a stunner. Mr. Adams was one of the longest-serving and most powerful politicians in Utah, a solidly Republican state, and had won earlier re-elections with little opposition"

Notice he didn't say "brain power", the power they need is the usual kind, which the existing players can get.

WSJ: "As AI Companies Race for Power, Amazon and Google Have the Lead.. Amazon has an incumbent advantage, and Google stands out for some innovative approaches.. Amazon’s chief executive has said that power is the 'single biggest constraint' in its cloud and artificial-intelligence business"

LLM is a one-tech-trick that Google had when it started, via Search. Google then started offering other IT services and enterprise components under its cloud umbrella. The chatbot business is the new one-tech-trick, OpenAI and Anthropic were among the first to develop it, they would attempt, in the future, to enter the cloud service biz just like Google had alongside their one-tech-trick. But the existing players will not let that happen. Google released its LLM freely, to crush them. There are open weight models that anyone can build on.

I don't think it matters. LLM research is a dead-end, you don't need anymore exploration in that area. It's time for commodatization and scale. Google can do both, with or without those guys.

"Top AI researchers Jonas Adler and Alexander Pritzel are leaving Google for Anthropic, according to Bloomberg."

"@GrapheneOS@grapheneos.social

Hyundai and Kia added official GrapheneOS support to their apps months before Volkswagen banned GrapheneOS:

Pressure from Volkswagen customers on them can achieve the same thing. There's no legitimate reason to ban GrapheneOS so they'll undo it with pressure."

MS Now: "Mamdani pushes Democrats to the left, one primary at a time"

Yahoo: "Pro-Israel politics just took a huge hit in New York.. Progressives who ran hard against the war in Gaza swept a trio of congressional primaries Tuesday in America's most heavily Jewish city, turbocharging the Democratic Party's yearslong shift away from Israel."

The F-16 vs F-22 (Raptor) win is not a 100% slam dunk. F-22 can shoot beyond visual range equipped with longer range radars and missiles, so F-16 can die before any dogfight starts. Maybe F-16 could defend against that via an upgrade kit that gives longer range capabilities. But out-of-the-box you'd have a problem.

There are other disadvantages. If you fly too slow, you die. You get into a one circle fight instead of two, you die.

How can F-16 make sure it remains in two circle? Apparently the second plane that turns makes that decision. So you wait, watch for the turn, then turn opposite.

"NASA satellite captures wave of warm water hundreds of miles long that signals a devastatingly strong El Niño"

The Guardian: "UK prioritised ties with UAE over averting mass atrocities in Sudan, MPs to be told"

TAC: "Markets Marred By Global Tech Rout.. Worries over an AI bubble and rate hikes helped to drive down tech stocks."

Sirota: "@ZohranKMamdani actively helping populist anti-establishment candidates in primaries is the opposite of what @BarackObama did 20 years ago.

The contrast reflects the differences between them - and shows how we're (finally) in a new political era.

This is from 20 years ago"

Apparently Charlie's Angels was a shitshow for its writing, they hired and fired writer after writer, everything was in chaos, but somehow it all came together in the end. Usually though changing the script too many times, going through lots of writers is a sign the production is doomed.

Link

Iconic scenes when good guy turns bad.. Norton in Primal Fear, or The Score?

Chollet argued "program synthesis" is the way to go for AGI. AI needs to learn "programs rather than representations of word clouds".

Are coding harnesses / agents are a way to get program synthesis via LLMs, through the backdoor? Maybe. Research can look into that, their equivalence if it's there, via a proof. But LLM's crooked world model will always get in the way even if the two were equivalent. We need progress on a few fronts... LLMs need to be consistently smart, on everything. We can't have the situation Jeremy Howard describes and power users frequently experience, 80% of the time "they are so smart, do so much for me" and 20% of the time "these things are f-ing dumb".

Hogg: "Hard to imaging a better way to spend a Tuesday night than watching a bunch of establishment, corporate & AIPAC backed Dems get destroyed. If this isn’t the Dem tea party I don’t know what is."

David Hogg: "Congress needs more Claires [Valdez, recently won the Democratic primary for New York's 7th Congressional District]: a renter, union organizer, and 36-year-old progressive who will defend working Americans, not corporations, real estate developers, or the special interests making our lives hell. Leaders We Deserve is beyond proud of our support.."

The Lever: "Midterm Madness: Mamdani’s Midas Touch.. Democratic socialists just doubled their numbers in Congress."

🤯 🤯 🤯 😂 😂 😂

Business Insider: "In a[n] interview.. Uber's operations chief, Andrew Macdonald, said it was becoming harder to justify AI costs within the company. He said that Uber CTO Praveen Neppalli Naga went viral after telling.. that Uber had already blown through its Claude Code budget for 2026. The comment led to what he described as a 'head-exploding moment,' sparking discussions about AI token consumption within the company and the trade-offs it creates, such as on head count. He said that, based on talks with Uber's senior engineering leaders, he realized higher token usage did not translate into a proportional increase in useful consumer features."

The Verge: "Uber president says AI spending is getting 'harder to justify'"

"@slava@mathstodon.xyz

Open source projects will tell you they have no choice but to allow slop, because doing anything else is a form of gatekeeping that will exclude too many potential contributors.

Nothing could be further than the truth, though. When the review queue is a non-stop machine gun of giant, bogus drive-by patches, the good contributions will inevitably fall through the cracks.

Eventually, the contributors you actually want—those who care about the long-term health of the code base, building institutional knowledge, improving their craft, mentorship—will simply leave, because they will realize their limited time and energy is better spent elsewhere.

A blanket ban on LLM contributions is not impractical, idealistic, or too radical—it's just common sense."

Another energy based approach, and looks probabilistic, there is sampling, albeit offloaded to the LLM

Paper: "Large Language Models (LLMs) struggle with reliable mathematical reasoning, and current verification methods are often computationally expensive. This paper introduces the Energy Outcome Reward Model (EORM), a highly efficient, lightweight post-hoc verifier designed to address this challenge. EORM uses an energy-based framework to rank Chain-of-Thought (CoT) solutions, learning to distinguish correct from incorrect reasoning using only simple outcome labels, thus eliminating the need for expensive annotations. With only 55M parameters, over 127 times smaller than typical reward models, EORM boosts the accuracy of Llama 3 8B to 90.7% on GSM8k and 63.7% on MATH."

[-]

Dario is lying his ass off about costs and revenue.. "Better earnings from more expensive model current year arrives next year bro so on annual basis we are always showing a looossss". Yeah, whatever.

"@LukaszOlejnik@mastodon.social

OpenAI shipped a telemetry system that logs more than the actual work being done. Codex burning through SSDs at a rate of ~640 TB/year – one user hit 37 TB written in 21 days. On a consumer SSD that’s full drive death in under a year."

#Cosplay #LLM #JeremyHoward

[-]

Reshare

#ProbabilisticGR #Barandes

[-]

The Guardian: "[2025] Universe expansion may be slowing, not accelerating, study suggests.. Astronomers cast doubt on Nobel prize-winning theory"

Bloomberg.com: "SpaceX’s High-Grade Debt Brings Out Skeptics.. S&P, which grades SpaceX one notch lower than Moody's at BBB, expects the company to remain cash-flow negative until 2030, with the burn rate rising sharply next year and again in 2028. To help finance that gap, SpaceX is expected to lean much more heavily on debt, with borrowings climbing to $132 billion in 2028"

Novara Media: "Gaza Genocide a Factor for Majority of Progressive Voters Abandoning Labour, New Polling Shows"

Jeremy Corbyn: "Keir Starmer could have ended child poverty, homelessness and the grotesque levels of inequality in this country.

Instead, he abandoned those in need, destroyed our civil liberties and facilitated genocide in Gaza.

That is how this Prime Minister will be remembered - and that is the legacy of moral and political bankruptcy he leaves behind.

The crises in our society are not going away. Neither are we - and we will keep fighting for a more equal, peaceful and dignified society for all."

You have to give it to the Brits: when MPs feel their future electability is in jeopardy due to an unpopular leader, they smack the living shit out of him (or her) and s/he is gone. After major Labor upset, Labor MPs picked the one candidate popular in his mayorality, and had a chance nationally, and they made it happen - enter Andy Burnham.

Burnham had to be in the parliament to be a PM, he sought to re-enter by applying to be the Labour candidate for a by-election, Starm blocked it. Then an MP "selflessly" resigned his safe seat in Makerfield to clear path for him, he achieved a landslide victory with 55% of the vote. Finally Starm saw the writing on the wall and "placing his country first" decided to resign.

Reshare

"The Surprising Flaws in 18650 Lithium-Ion Batteries"

[-]

#Karpathy #Vibe

[-]

Reuters: "UN chief calls on AI firms to come clean on environmental costs"

BBC: "Mamdani-backed candidates win in New York's Democratic primary"

Reuters: "US Senate joins House in voting to halt Iran war, rebuking Trump"

Commodatization. Soon companies will not be able to make money simply by being an "AI company". Once cloud IT service firms start offering free LLMs that are easily installed by choosing it from a dropdown, the business model of said "AI companies with frontier models" will experience revenue loss.

Business Insider: "What is GLM-5.2? Another open-source Chinese AI model has Silicon Valley's attention."

Some in US use the term to mean "bureucracy". The Anglo is so brain mucked by free marketism anything non-free market seems like "deep something bro" to them... They are fleeced by the people in front of their eyes, who feed them candies, sell them crooked products, but they don't notice. I guess it's like the saying "a fish does not know what water is".

Deep state is used for dark corners of bureucracy that elected politicians cannot control, not the whole bureucracy. In US politicians can control the bureucracy esp. on security matters.

US has no deep state.

The term "deep state" was invented in Anatolia (so-called Turkey), so we own its definition.

I bet he consulted Greenspan on that one too, and he said "go for it Bill"

That literally sounded like Ronald Reagan

Treasury.gov: "[1999] Statement by President Bill Clinton at the signing of the financial modernization bill.. 'We are here today to repeal Glass-Steagall because we have learned that government is not the answer'"

Basically Clinton extended the reign of the most economic right-wing character you can imagine, and a Republican to boot, well into the mid 2000s, until a few years before the Great Financial Crisis btw.

Wiki: "Democratic president Bill Clinton reappointed Greenspan, and consulted him on economic matters."

NPR: "One of the most important intellectual relationships in the life of Alan Greenspan.. was with author Ayn Rand, whose 1957 novel Atlas Shrugged has become a perennial favorite among conservatives and which the Library of Congress named as one of the books that has shaped America."

The man responsible for running the economy did not grasp the most fundamental aspect of economic behavior. Is that not scandalous? How do the people sharing the same ideology with Greenspan still remain libertards? King libertard said he messed up because of his ideology.

HE WAS SHOCKED

Reuters: "In addition to critiques of his monetary policy, critics slammed Greenspan, a powerful advocate for the light regulation of financial markets, for a hands-off attitude that allowed banks to make disastrous housing market bets... Greenspan subsequently admitted to being 'shocked' that he was wrong in his assumption that bankers' self-interest would deter them from taking actions that imperiled the survival of their own institutions."

That jagoff was a libertard btw, Ayn Rand style

WaPo: "Alan Greenspan, most powerful central banker of modern times, dies at 100.. His nearly two-decade run as chairman of the Federal Reserve helped spur prosperity, but his decisions also contributed to the 2008 financial crisis."

An expert claims the layoffs are still about offsetting the massive build up after covid when money was cheap, not replacing workers with tech. The replacement narrative is used to bolster company's image to look cool "in the age of AI".

Fast Company: "Meta CTO: Company morale is near the ‘worst it’s ever been’ after layoffs.. After 10% of Meta’s workforce was laid off, Andrew Bosworth said the tech giant is working toward improving employee morale, according to reports from Business Insider and Wired."

IC: "Trump Is Forcing Coal Pollution on Consumers and Communities.. The president's unilateral moves on coal are exposing more Americans to toxic mercury pollution"

Poetic justice.. you create chaos around the world along with your "big brother", and chaos finds you.

Larry the Cat has seen off six already, he is about to meet a seventh.

"@jalefkowit@hachyderm.io

Soon there will be enough living former UK prime ministers running around to field a soccer team"

IC: "In Gaza, Fathers Can’t Promise Their Children Food or Even Survival"

"@GeofCox@climatejustice.social

Let's not lose sight of this: Starmer failed in the way that all centre-left governments are doomed to fail, because there are no non-radical solutions to the polycrisis. The centre-left dream of positive change without disruption is over."

BBC: "Keir Starmer resigns in emotional Downing Street speech"

Reshare

Paper: "Boltzmann-GPT: Bridging Energy-Based World Models and Language Generation.. Large Language Models (LLMs) have emerged as remarkably capable language generators.. However, a fundamental question remains: do these models understand the world, or do they merely generate plausible texts about it?..

This concern has gained attention, with LeCun describing autoregressive LLMs as 'doomed' due to their inability to model causality, and Hassabis emphasizing that artificial general intelligence (AGI) requires world models capable of understanding physical reality. LLMs learn to talk about the world, but whether they have formed a coherent model of it is far from clear. The mouth speaks fluently; whether the brain comprehends is another matter. This motivates our architectural principle: the mouth is not the brain.

To rigorously evaluate this architecture, we use deliberately minimal components: a Deep Boltzmann Machine.. as the world model and a frozen GPT-2 as the language model. This choice is intentional. Frontier LLMs may have already internalized implicit world models, making it impossible to disentangle the contribution of explicit world modeling from the LLM’s own capabilities. By pairing a language model with limited capacity against an explicit energy-based world model, we can cleanly isolate the causal role of world model conditioning, ensuring that performance gains stem from structured understanding rather than the language model’s latent knowledge."

[-]

"The Surprising Flaws in 18650 Lithium-Ion Batteries"

[-]

BBC News: "Why lithium‑ion battery fires are so dangerous"

[-]

Some on the list would have directly benefited from the 2008 bailout

BMO Financial Group
Capital One Services LLC
Fifth Third Chicagoland Foundation
Northern Trust
Prudential Financial, Inc
GCM Grosvenor
Fidelity Charitable / National Philanthropic Trust:
Tony and Amie James (longtime President and COO of The Blackstone Group)
Catherine M. and Frederick H. Waddell
Jonathan and Jeannie Lavine

The Lever: "With his presidential library, former President Obama is unveiling an oligarch-funded shrine to himself amid everyone in America being fleeced by oligarchs.. [The library's] sponsor list is context for a presidency that promised hope and change and then used a massive electoral mandate to deliver more of the same. Indeed, the [donor] list is a who's who of the winners of the Obama era: tech moguls, financial giants, telecom behemots, a health insurance giant, and other bold-faced names of the oligarchy"

"Clinton's Presidential Library Raised 10% of Funds Overseas"

ABC News: "[2001] The ex-wife of fugitive commodities giant Marc Rich donated $450,000 to the Clinton presidential library before then-President Clinton pardoned the billionaire last month"

#Beato #LLM

[-]

"@ainmosni@ainmosni.eu

We need more places to openly take stands like Thomas House."

[-]

"@LukaszOlejnik@mastodon.social

Claude Code AI coding assistant is tracking its users & collecting lots of data. Did you know of this? Default-on behavioral metrics to Anthropic every 5 minutes."

Presence of constants BTW point to a lack of knowledge in that area, not presence of it. The so-called Standard Model of physics has 26, and its adherents are stuck in the mud, they cannot move an inch further improving the theory.

In the previous post we said "LBM relaxation toward the equilibrium can be seen as N-S viscosity". That should not be seen as validation, as if the analytical approach has primacy and if simulation matched that it is more correct.

And, what is viscosity anyway, the $\mu$ in Navier-Stokes. 90% of the time it is assumed constant (makes sense, consistency, thickness of a fluid does not vary in most conditions). So the formula says there are gazillion of inter-molecular interactions, and that whole dynamic, via friction, other cohesive forces, results in one stupid constant? So IOW the analytical also summarizes, it generalizes, it is not the real thing.

Don't get me started with the "in the limit" arguments.. In the limit, distances approach to the infinitesimal. But inter-molecular distances, albeit small compared to regular objects in the world, are not infinitesimal. So the math is essentially wrong (though useful).

I wonder what would it be like discovering Lattice Gas Automata and later its descendant, Lattice Boltzmann Method before the analytical formula Navier-Stokes? Since both are equally fundamental it should be okay to invent the simulation formulation before the analytical approach (the grand formula).

#ProbabilisticGR #Barandes #TOE

[-]

Carry-On, fantastic work. I like who the good guys and the bad guys are.

The Guardian: "[2008] James Lovelock: 'Enjoy life while you can: in 20 years global warming will hit the fan'.. The climate science maverick believes catastrophe is inevitable, carbon offsetting is a joke and ethical living a scam."

"@Nigel_Purchase@mstdn.social

France's hottest day ever"

LLM's (book) smarts converges to the average, widely-known, so they can be used for ed assistance. If the LLM is not in the driver's seat, only used as a helper while reading an authoritative book on a subject, they can be ok. If there are unclear points in the book, student could ask LLM "how is formula B derived from formula A?". LLM will go through its neural spaghetti and spit out something likely on how that derivation was achieved. This is key: the process needs to start and end in the authoritative material. Then student can judge LLM's explanation, if its internal logic makes sense, the conclusion leads to exactly what is in the authoritative source, accepts / rejects based on that result. LLM is basically used to fill in the blanks. Make the LLM your bitch.

#Ukraine 06/10 - 06/21

[-]

"@GeofCox@climatejustice.social

The Israel/US regime-change assassination strategy backfired catastrophically, bringing to power in Iran new people willing to assert the control over the whole region, and influence over the whole world economy, that was always latent in Iran's size and geography."

"@Gizmodo@flipboard.com

Waymo Recalls Over 3,800 Robotaxis Over Risk of Driving Into Freeway Construction Zones"

Deep Water, ok movie.. It started great, lost some of its steam w/ B-level writing here and there in later acts... it is also clear some Chinese money is behind the production.. You can detect the propaganda. The obnoxious American guy could literally play the white evil antagonist (gweilo) in Hong Kong cinema (maybe he has). But hey can we blame the Chinese for doing little media campaign of their own? If Ben Kingsley, Aaron Eckhart liked the script thats good enough for me.

"@pikesley@mastodon.me.uk

From the perspective of the Xenomorph, Alien is basically Die Hard"

The technique mentioned in the paper is a true energy based model / probabilistic. They use (Deep) Boltzmann Machines for their world model that are stochastic neural networks. Good to hear DBMs are still alive and being used in new research.

[-]

#TheAmericanDream

Link

The movie Pressure has Brandon Fraser as Ike but why is he fat like Churchill?

"Mia Mercado at The Cut delved into the culinary frontier of AI-generated recipes, and in the several-course-meal of this endeavor, subjected herself to eating 'literal AI slop' that she cooked herself, to see if the instructions held up in reality...

On TikTok, she found a recipe video for cottage cheese breadsticks which was entirely AI-made, down to the voiceover, the loud 'crunch' of the food, and the physics-defying visuals. 'After combining the blended cottage cheese with an egg and mozzarella, I added the optional (???) garlic powder, salt, and pepper. The batter was loose, and my hopes were low,' Mercado wrote. 'How would this pan of goop become twisted, bready sticks?'

Narrator: they didn’t.

'It’s more like an eggy sheet with herbs,' Mercado lamented. 'When I tried to twist the strips to resemble the original video, three of them fell apart. They tasted like a weird omelet.' Their final rating: '0 out of 5 AI-generated thumbs'"

Fortune: "OpenAI’s financials have leaked, showing $21 billion in losses against 13 billion in revenue"

"@jhpot@mastodon.social

Saw the CEO of my former employee talking about his council of AI advisors on LinkedIn and am sincerely wondering if I should reply with links to mental health resources"

"The Earth Prize 2026: Indian teen trio named Global Winners for magnetic tamarind powder that removes microplastics from water.. The Earth Prize is the world’s largest environmental competition and ‘ideas incubator’ for 13-19 year-olds, empowering young people with mentorship and shared $100K funding.. Global Winners team Plas-Stick from India created a magnetic tamarind powder that removes microplastics from water"

At the level of large orgs, software isn't just about "writing code" - it is about orgs having someone responsible for the code delivered. We talked about how many enterprise clients will not even accept freely available code unless someone is legally liable for it. Same will apply to vibe coded apps. Who is responsible if something goes wrong? Who can you sue if code breaks? Anthropic, OpenAI sure for shit won't take any responsibility for the kind of spaghetti "neural" code they produce. That leaves consultancies, or other coding service providers -just like today- they might use LLM to code part of their codebases, but will eventually be personally on the hook for the correct execution of the program they deliver.

For demo code, "write-only" type of scripts, producing something for yourself, fine. But enterprise level coding will still require professional programmers. Hugh Jass from sales won't be able to vibe code for his department all on his own just because that code cannot completely be trusted.

Of course they can.

Wired: "Can normies really vibe code?"

Captain America: Corporate Soldier

[-]

"AI or DIE TRYING"