Posts by O'Reg Inalsin

Power: The answer to and source of all your AI datacenter problems

Re: If power and infra are the true bottlenecks

Quite so. However this statement - the west thinks 1-4 quarters requires a deeper look.

Firstly, financial returns of some sort over the next 1-4 quarters are indeed actually important, anywhere in the world, but those returns are increasingly derived from changes in stock prices and are increasingly speculative. There is a reasonable argument that this is a good thing in that the stock purchases are like long term loans that appreciate as the company gets closer to the date when profitability is achieved and the stocks start paying dividends. I.e., it enables looking 10 years ahead.

However a closer look reveals that the long term net rise in the stock market has been assisted by repeated government intervention (fiddling with bonds, reserve requirements, QE, psuedo-QE) and the rise of ETFs, which makes the whole thing too big to fail. There is a reasonable argument that this is a bad thing, because it is the antithesis of capitalism that keeps the beast alive with government life support, and has enabled monopolies or generally the largest players, K-street lobbyists, and the financial industry (now conjoined with bitcoin) to thrive at the expense of smaller businesses. It has also kept the dollar high, encouraging outsourcing, and hurting manufacturing exports - which have proven to be problematic.

So what does this have to with rapid expansion of Datacenters and online computing power? There is an argument that this is a modern form of manufacturing - making tokens - which is a form of productivity. This argument goes that this can be the US' answer to China's 50 years of government subsided physical exports and domination of manufacturing. In 50 years the world will rely on (exponentially closer to) "AGI" tokens the same way that the US now relies on Chinese manufactured goods. Except that it won't be US government supported but supported by justified speculative stock prices.

I've run out of time but that just won't work.

Re: If power and infra are the true bottlenecks

That was the cheapshot pun!

If power and infra are the true bottlenecks

... then AI supply must be outstripping AI demand and OpenAI is making money hand over fist by jacking up prices to profit, profit, profit. Strange ... is something is not adding up?

Meanwhile China is spending less on datacenters but more on manufacturing automation and other quick to profit applications of AI.

In short, with regards to AI, China appears to be overtaking the US on following capitalist principles with respect to AI investment, while the US is 100% focused on creating an almighty God to be mediated with through annointed human holy prophits, all in the name of the Holy AI Crusades.

Memory boom-bust cycle booms again as Samsung reportedly jacks memory prices 60%

Re: Blame it on AI

Yeah, sure, but Nadella recently said in an interview - "So, if you can’t do that [have the power and datacenters ready], you may actually have a bunch of chips sitting in inventory that I can’t plug in. In fact, that is my problem today. It’s not a supply issue of chips; it’s actually the fact that I don’t have warm shells to plug into.” Some interpret this as meaning MS is stockpiling GPUs and related hardware, which would be exacerbating prices.

MS CFO Amy hood previously said “We are, and have been, short now for many quarters. I thought we were going to catch up. We are not. Demand is increasing,” So why not raise prices and make a profit? Using pole position to undercut any possible competition. No AI specifically, but abuse of AI certainly.

Software engineer reveals the dirty little secret about AI coding assistants: They don't save much time

Re: The student becomes the master.

AI can never be anything more that a tool. But AI can be a tool that enables the user to become someone (or something) else's tool. To be fair "being a cog in the wheel" is not always a bad thing - being a member of a team following a leader or an overarching abstract ideal, can be very satisfying and productive, as one part of one's life, in a symbiotic relationship with one's own human originality. In the worst case however, it kills originality of action and thought - Dark Ages stuff. AI has the power to let us lead us into the new Dark Ages, if we let it.

Ransomed CTO falls on sword, refuses to pay extortion demand

Re: Chief Technology Officer Mariano Albera said that his company takes "full responsibility"

I understand, but there is no shortage of wealthy people and powerful organizations that literally steal from people with less resources then them.

Big Tech's control freak era is breaking itself apart

Re: This article was timely

Plan A:

- use the current billions in hand to buy up all the GPUs, memory, ssds, etc. to kill all competitors and non-AI-cloud usage. (Already underway)

- outsource all computing work. "Temporarily", because AI "is not yet" capable of replacing humans, wait for AGI "next year. (Already underway)

- Feds spend 1 trillion on AI cloud usage - NOT a bailout. obviously - just "Efficiency" (Coming soon)

Here's one way to cut support ticket volume… send them to another company entirely

Re: Doesn't have to ge AI

Automated stupidity is superior.

Senate bill would require companies to report AI layoffs as job cuts reach 20-year high in October

The Firing Squad

WSJ - "The Boss Has a Message: Use AI or You’re Fired"

At enterprise-software company IgniteTech, leaders required staff last year to devote 20% of their workweek to experimenting with AI. On one such “AI Monday,” staff brainstormed ways to speed up processes like automating responding to customer-service tickets. Employees also had to share on Slack and X what they were learning about AI.

CEO Eric Vaughan said that employees self-assessed their AI usage and, afterward, the company used ChatGPT to rank the results. After a human review, IgniteTech cut the lowest-scoring performers.

“By their own admission, they’re in the basement,” he said. “So now they have to leave.”

It wasn’t easy: Vaughan recalls speaking with his wife over that time about the changes, feeling “terrible.” But he said he felt AI was an existential threat, and that if IgniteTech didn’t transform, the company would die. One tough exit was the chief product officer, who had been with the company for years. He and others were model, productive employees historically but were resisting the AI mandate, said Vaughan, who also leads GFI Software and Khoros.

Greg Coyle, that executive, said he had bought into AI’s potential to improve IgniteTech’s products and add new capabilities. But he took issue with the nature of the widespread cuts, particularly because the technology is in such an early stage.

“Doing this rapid culling of your workforce, it’s very risky,” he said. “If your AI plan doesn’t work out the way you expected it to, it’s a huge risk for the business.”

After a round of cuts, Coyle said he pushed back against an AI mandate in late 2023 in an executive meeting. He said he felt the company wasn’t working strategically as it pushed out staff. A few months later, he said, he was fired.

AI, Coyle said, is “coming whether we like it or not. You either get on board or you get left behind.” But, he added, “I don’t believe that you take this brute force, across-the-board approach to AI in the business.”

Google’s Ironwood TPUs represent a bigger threat than Nvidia would have you believe

Can we dream? - Autonomous AI agents capable of making high return bets in the stock market

From WSJ, "AI Is Co-Writing Financial Reports. Here’s Why That Matters." --

... Another company, Hewlett Packard Enterprise, which has used its large language model for earnings-day prep, is getting ready to use it to formally produce a first draft of financial statements, possibly for the quarter ending in January, said Chief Financial Officer Marie Myers.

Researchers want to kill the vibe, propose better model for AI coding

Re: ...like a book

While a fiction book is read linearly, good design of software is usually best described in tree form (*) with top down design and bottom up implementation. (*Some cross connections too, but as few as possible). Some books which are not fiction or entertainment oriented also have a tree-ish format - such as manuals.

AI's trillion dollar deal wheel bubbling around Nvidia, OpenAI

As reported by Tom's Hardware, Satya Nadella said in a recent interview - “... you may actually have a bunch of chips sitting in inventory that I can’t plug in. In fact, that is my problem today. It’s not a supply issue of chips; it’s actually the fact that I don’t have warm shells [data centers] to plug into.”

This pretty much sounds like MS buying up memory and GPUs that aren't being used just to prevent any competition from developing. Including domestic competition.

From Intel to the infinite, Pat Gelsinger wants Christian AI to change the world

Re: Bring out the comfy chair!

Some quotes from the 1825 Book by LaPlace - "Essai Philosophique sur les Probabilités", chapter - "Des illusions dans l’estimation des probabilités"

"One of the great advantages of the probability calculus is that it teaches us to distrust our first impressions. As we discover, when we are able to submit them to the calculus, that they are often deceptive, we ought to conclude that it is only with extreme circumspection that we can trust ourselves in other matters."

"Finally we shall establish, as a psychological principle, the exaggeration of probabilities by the passions. Something that is feared, or that is keenly desired, seems to us for that reason to be even more probable. Its image, strongly etched on the sensorium, weakens the impression of contrary probabilities, and sometimes obliterates them to the point of making one believe that the thing has happened."

What LaPlace quote are you mentioning?

Re: Bring out the comfy chair!

I do not see how science and religion could be unified, or even synthesized, under any common scheme of explanation or analysis; but I also do not understand why the two enterprises should experience any conflict. Science tries to document the factual character of the natural world. Religion operates in the equally important, but utterly different, realm of human purposes, meanings and values. I propose that we encapsulate a central principle of respectful noninterference - accompanied by intense dialogue between the two distinct subjects - by enunciating the principle of Noma, or non-overlapping magisteria (from the Latin magister, or teacher). Magisterium is, admittedly, a four-bit word, but I find the term beautifully appropriate. To summarise, the magisterium of science covers the empirical realm: what is the universe made of (fact) and why does it work this way (theory). The magisterium of religion extends over questions of ultimate meaning and moral value. To cite the old cliché, science gets the age of rocks, and religion the rock of ages. - Stephen Jay Gould, 1999, "Rocks of Ages: Science and Religion in the Fullness of Life"

Not a completely new idea - e.g., Einstein said something even better - "Science without religion is lame, religion without science is blind.” - 1941 essay "Science and Religion".

We biological Earth beings evolved through hell - literally fire and brimstone. That school of hard knocks left us (animals) with a shared complex web of emotions and instincts. For we humans, religion began as the way to make sense and impose control on those internal forces emanating from that legacy.

AI can never personally experience the forces of that volcanic legacy, which is something that apostate "AI as the superior being" evangelists completely fail to understand. I sincerely hope Gelsinger realizes he is tiptoeing around volcanic hot springs here.

Canonical CEO says no to IPO in current volatile market

Canonical is now focused on Ubuntu plus other software in containers for the cloud.

Canonical has shifted emphasis from desktop to cloud containerization, which is where they are making money. I quit Ubuntu desktop after they closed the backdoor to allow installation on pre-encrypted disk - you now have to tell the key to the Ubuntu installer and let it encrypt the disk for you. Also, they no longer offer packaging support for the entire "supported" duration of a release - after a certain date if you want to install "tree", for example, you are instructed to install the tree snap package. Canonical no longer really care about desktop, and anyway, Debian now "just works".

Invisible npm malware pulls a disappearing act – then nicks your tokens

Re: This is a bug in npmjs.org

Bleeping computer> "The mechanism fetches packages and executes them automatically when running ‘npm install’, and requires no user interaction."

So it seems! Yikes. This is on the package manager, isn't it? I suppose they are all promiscuous like that.

I wonder if DENO has taken a different approach.

Re: This is a bug in npmjs.org

I'm confused. The article stated that the packages hasd "no dependencies", but your explanation is clearly a dependency of a type that could be easily detected (explicit unknown url). Does this vuln depend on automatic version upgrades which introduces depencies in a way that evades scanning?

The article also states the execution takes place during the installation, rather than when the software is executed by the user. Package managers don't purposely run the software, I think, so there is a step I don't understand.

Trump and Xi ease trade tensions, but Nvidia still can't sell Blackwell in China

Re: The USA can do without Canada

In 2024, Canada was the largest purchaser of U.S. goods exports, totaling approximately $356.5 billion, while exports to China amounted to about $150.4 billion. China is a huge market, but the barriers have always been high. The drop from x3 to x2.3 difference in 2024 was perceived by China as a belligerent offense.

Smile! Uncle Sam wants to scan your face on the way in – and out

The problem here is that non US visitors have recently become assumed guilty until proven innocent, to a higher degree than before. Having your photo taken is nothing that new, especially compared to the UK.

Ex-CISA head thinks AI might fix code so fast we won't need security teams

Re: What an idiot

"We don't have a cybersecurity problem. We have a software quality problem," she said. The main reason for this was software vendors' prioritization of speed to market and reducing cost over safety.

... the real focus should be on the fact that the common factors uncovered by MITRE nearly 20 years ago – cross-site scripting, memory unsafe coding, SQL injection, directory traversal – remain part and parcel of shipped software. "It's not jaw dropping innovation… They were the golden oldies."

So far so good. However ...

This is because software companies insisted customers bear all risk and convinced government and regulators that this was acceptable.

That's far too broad of a statement. Suppose instead a NIST public standard such as "this software contains no listed CVE's as of MM/DD/YY". If a realistic standard existed public companies would effectively have to require it to meet their own standards. But any such realistic standard(s) would not be a cure all.

Finally, the kicker:

AI offers a way to address this, she claimed, as it is far better at tracking and identifying flaws in code. And it would be possible to tackle the mountain of technical debt left by a "rickety mess of overly patched, flawed infrastructure."

Here Easterly is paraphrased as saying "it [AI] is far better at tracking and identifying flaws in code", which is at least very vague - better that what?. AI is a tool that can assist humans, not an intelligence that can replace humans, unless you want to introduce even more convoluted vulns. Perhaps the reporter's paraphrasing of what Easterly said was unfair? It's this last paraphrased statement that has really ticked Reg readers. Probably she she deserves a chance to respond to this paraphrasing before being condemned in the court of El Reg.

Signal president Meredith Whittaker says they had no choice but to use AWS, and that's a problem

Re: Depends on their use case specifically

To me, one of the best (on paper) use cases for public cloud is you have to go from say ~100 CPU cores to 5,000 CPU cores for max of 2 hours per day ...

In the days of new internet startups startups, in particular social media, waiting for that lucky viral moment when usage would suddenly explode was key to getting an IPO, and missing such an opportunity was like missing the boat. That whole industry though doesn't produce anything worthwhile except crappy LLM training data, and negative cultural impact of viral+shallow is horrifying.

Re: Really?

We had a lot of people that focused on resume driven development ... - Does this mean employees, presumably management with such decision making authority, directed services to be ported to AWS because it looks good on their resumes?

MPs urge government to stop Britain's phone theft wave through tech

Don't let the choke be on you

Make sure you get a lanyard with a quick release that will disconnect under slight tension - it won't save your phone but it will save your vagus nerve and spinal cord.

AI investment is the only thing keeping the US out of recession

AI investment vs. AI bubble betting

I would argue that "AI Investment" needs to be distinguished from "AI Bubble Betting", and also that "AI Bubble Betting" is aimed at preventing investment (including other AI investment) outside of a very narrow circle of companies. It is inherently anti-competitive, aiming ultimately for monopoly, or near monopoly, domination, in coordination with lobbying contributions for political cover. This is far closer to the pre-capitalist economic model (of a king delegating licenses to operate monopolies in exchange for royal patronage) than a system that ensures a level playing field to reap the benefits of competition.

The massive AI-scaling buildout behind the current "AI Bubble Betting" is actually contributing to a recession by taking resources away from the rest of the economy (including other AI investment).

A single DNS race condition brought Amazon's cloud empire to its knees

Re: Looks like they still didn't catch the cause

Async computing 101, atomic modifications. The same category of whoopsie that triggered the British Post Office Limited catastrophe.

New Linux kernel patch lets you cancel hibernation mid-process

Dilemma

If checked baggage with battery inside gets used for shotput practice, then put in the hold, how long does it take to detect the fire compared to abused baggage in an overhead bin?

Trust the AI, says new coding manifesto by Kim and Yegge

They can happily vibe code to their hearts content because they can spot when the AI gets it wrong.

I think it likely both of them are are in the business of building and debugging AI software tools, and therefore - as they should be - eating the dogfood as they bake it.

Which is fine! But it also means that the book is not an unbiased source, and that their enthusiastic usage is not the proof of a finished safe and productive tool they would have you believe.

SpaceX is behind schedule, so NASA will open Artemis III contract to competition

Sampling the goods is a no no

That Ketamine was mean to be used only as a rocket fuel additive, not for personal consumption.

Amazon brain drain finally sent AWS down the spout

The more things change .... (Gilbert and Sullivan's 1879 comic opera The Pirates of Penzance)

I am the very model of a modern Major-Gineral

I've information vegetable, animal, and mineral,

I know the kings of England, and I quote the fights historical

From Marathon to Waterloo, in order categorical;

I'm very well acquainted, too, with matters mathematical,

I understand equations, both the simple and quadratical,

About binomial theorem I'm teeming with a lot o' news,

With many cheerful facts about the square of the hypotenuse.

...

For my military knowledge, though I'm plucky and adventury,

Has only been brought down to the beginning of the century;

But still, in matters vegetable, animal, and mineral,

I am the very model of a modern Major-Gineral.

US hyperscalers to guzzle 22% more grid juice by end of 2025

Re: Ha!....The Place To Be Is China.....

Narrowly correct - narrowly meaning Krugman ignores many other factors, e.g. that Chinas coal energy generating capacity in 2024 was 1150 GW compared to US 60 GW, and China is bringing on about 80~100 of new coal powered GW online in 2025 - more than the entire US coal generating capacity.

Literal crossed wires sent cops after innocent neighbors in child abuse case

Re: They should have sued BT

The police force could and should sue BT for wasting their time and interfering in a criminal investigation by negligence.

Re: How were children removed when search turned up nothing?

Clearly every RIPA execution needs to be accompanied by a BT employee checking that the line is going to the address they think it is going to.

How were children removed when search turned up nothing?

The trio brought the case as they argued their Article 8 rights (right to a private family life) under the RIPA were infringed. These claims largely hinged on the idea that the police's RIPA requests for communications data from BT were unlawful because they could have made other lines of inquiry before issuing them.

This doesn't seem like the right track to claim damages. Considering the all three suffered actions amounting to being assumed guilty (employer informed, job offer lost, CHILDREN REMOVED) - shouldn't they be suing for real damages from those actions? The police could have made the RIPA request and made their two searches and seizures without executing those injurious actions on those people.

Chinese gang used ArcGIS as a backdoor for a year – and no one noticed

Re: Pots and kettles

The bigger danger is the drip drip drip of El Reg's overwhelming charm chipping away at VoT's hard resolve and willpower.

Mozilla is recruiting beta testers for a free, baked-in Firefox VPN

Re: Why spend money on building a [free] service and supporting its infra that nobody asked for?

"We collect only the technical data needed to keep Firefox VPN reliable and secure."

Just don't forget It costs money to keep Firefox VPN reliable and secure. Wonder why this isn't being released in GDPR-land?

Datacenter water use? California governor says don't ask, don't tell

Re: California water politics is already a corrupt oligarchic kleptocracy

CA produces 1/3 of the vegetables, 3/4 of the fruits and nuts. over 9/10 of the avocados, broccoli, lettuce, grapes, and strawberries, of all such products produced in the US. All those city dwellers require food - without food the GDP would go to zero pretty quickly.

As for the long term trend into larger and larger corporate farms, you can blame decades of de facto shadow-legal illegal immigration of farm workers, which obviously pushes the advantage to large scale farmers - 1 farmer employing 1000 undocumented workers is going achieve lower costs than 100 farmers employing 10 undocumented workers each.

Benioff retreats from idea of sending troops in to clean up San Francisco

Re: SF's homelessness problem is a housing problem

Contributing factor, yes. Sole reason, no.

Re: Salesforce has to pay for “hundreds of off-duty law enforcement officers”

From "Bandana Resources: How Much Do SFPD Officers Make?", which is recruiting for the SFPD: As an SFPD recruit in the academy, you start with an annual salary of $112,398, equating to approximately $53.89 per hour. This competitive starting pay makes SFPD one of the highest-paying law enforcement agencies in the Bay Area, ensuring you’re financially supported from day one. ... Salaries for SFPD officers increase significantly with experience and tenure. After seven years of service, officers earn $164,164 per year, which is approximately $68.65 per hour. With average overtime (7 hours a week), officers can expect total compensation of around $201,663 annually. For the top 10% of earners, total compensation, including overtime, can exceed $288,887 per year.

Gartner warns agentic AI startups: Prepare to be consolidated

Shock!

... while the winners would be capital-rich incumbents

McKinsey wonders how to sell AI apps with no measurable benefits

Alternative theory

Software vendors keen to monetize AI should tread cautiously, since they risk inflating costs for their customers without delivering any promised benefits such as reducing employee head count after firing the valuable employees they need to integrate use of AI at the ground level.

To the extent that AI is useful, AI is useful because it is a tool in skilled human hands.

All you need is clients looking for rubber "obfuscating" stamps, they won't care. For example the Delloite report with reviewed Australian departmental IT systems’ erroneus use of automated penalties in Australia’s welfare system:

Deloitte Australia will partially refund the 440,000 Australian dollars ($290,000) paid by the Australian government for a report that was littered with apparent AI-generated errors, including a fabricated quote from a federal court judgment and references to nonexistent academic research papers. The financial services firm’s report to the Department of Employment and Workplace Relations was originally published on the department’s website in July. A revised version was published Friday after Chris Rudge, a Sydney University researcher of health and welfare law, said he alerted the media that the report was “full of fabricated references.” “Deloitte had agreed to repay the final instalment under its contract,” the department said. The amount will be made public after the refund is reimbursed.

The first error that jumped out at him wrongly stated that Lisa Burton Crawford, a Sydney University professor of public and constitutional law, had written a nonexistent book with a title suggesting it was outside her field of expertise. “I instantaneously knew it was either hallucinated by AI or the world’s best kept secret because I’d never heard of the book and it sounded preposterous,” Rudge said. Work by his academic colleagues had been used as “tokens of legitimacy,” cited by the report’s authors but not read, Rudge said, addding that he considered misquoting a judge was a more serious error in a report that was effectively an audit of the department’s legal compliance. “They’ve totally misquoted a court case then made up a quotation from a judge and I thought, well hang on: that’s actually a bit bigger than academics’ egos. That’s about misstating the law to the Australian government in a report that they rely on. So I thought it was important to stand up for diligence,” Rudge said.

Employees regularly paste company secrets into ChatGPT

Re: circumvent our policies to access chatbots

Non they can add non-approved use of chatbot AI to that category.

Re: circumvent our policies to access chatbots

There are browser extensions known as "proxy extensions" and corresponding "proxy services" that forward https requests to bypass firewalls, similar to a VPN but just for one browser tab.

And of corse there is VPN also.

So an "ironclad" internal firewall still has gaps.

Bank of England smells hint of dotcom bubble 2.0 in AI froth

For some applications, hallucinations don't actually affect existing quality

Oct. 7 - Deloitte will begin using AnthropicAI's Claude assistant for hundreds of thousands of its global employees. - The global professional service giant Deloitte announced Monday that more than 470,000 of its workers around the world will gain access to Anthropic's artificial intelligence assistant Claude in San Francisco-based Anthropic's biggest deployment to date.

Oct. 7 - Deloitte’s member firm in Australia will pay the government a partial refund for a $290,000 report that contained alleged AI-generated errors, including references to non-existent academic research papers and a fabricated quote from a federal court judgment. The report was originally published on the Australian government’s Department of Employment and Workplace Relations website in July. A revised version was quietly published on Friday after Sydney University researcher of health and welfare law Chris Rudge said he alerted media outlets that the report was “full of fabricated references. Deloitte reviewed the 237-page report and “confirmed some footnotes and references were incorrect,” the department said in a statement Tuesday.

Cerebras CEO insists dinner-plate-sized chip startup will still go public

Amdahl revisited

That’s interesting: Gene Amdahl also attempted to develop wafer-scale integration (WSI) for supercomputing, founding Trilogy Systems in 1980. Despite raising an estimated $230 million from investors like Groupe Bull, Sperry Rand, and DEC, the project was plagued by disasters (including floods that delayed plant construction and ruined clean rooms). Ultimately, after burning through about a third of the capital, Amdahl concluded that success would require a 99.99% chip, unlikely for at least 100 years.

Today, Cerebras claims to have cracked the yield problem by designing chips with tiny, redundant processing cores and dynamic on-chip fabric reconfiguration to bypass defects. According to Cerebras, this allows them to utilize about 93% of the wafer surface area.

This El Reg article notes that Cerebras is also planning to undercut the competition, offering Llama3-70B inference at $0.60, plus a pinch of salt, per million tokens, compared to $2.90 on H100-based clouds.

It’s a humongous task. Not just the original technology, but also designing a GPU is enormously difficult (even just an LLM-specific-PU). But compared to building a working quantum computer, Cerebras’s target seems much more feasible. Good luck to them. I hope their momentum isn’t derailed by any unexpected AI bubble collapses. If I recall correctly, the tortoise wins the race.

Back to the future, predestined for success

In August 2022, Cerebras was honored by the Computer History Museum in Mountain View, California. The museum added to its permanent collection and unveiled a new display featuring the WSE-2—the biggest computer chip made so far—marking an "epochal" achievement in the history of fabricating transistors as an integrated part.

India's tech talent pipeline is sputtering

Re: Hang on

Instead of applying human talent to hands on develop and bring the tech to the office or factory floor, and expand in line with rising profits, there is more short term skimming in building out oodles of data centers that will never be utilized.