Posts by Fido

Zuck’s Chinese agentic prey escapes, will resume standalone ops

Alternative in Visual Basic

Giving open weights models trained at a loss away for free is called dumping. Dumping does not contradict keeping control of local innovators since there is no way to reverse engineer how the weights were created.

While chatbots are well known and as fun as a poke in the eye, how to make AI agents safe rather than dangerous and destructive is unsolved. The Manus AI agent represents regular code created using standard engineering practices. This code presumably has value as an understandable product of hard lessons and new ideas.

Since the 2 billion saved by not buying Manus is boring if left unspent, I suggest buying a professional sports team like all the other oligarchs. Alternatively, give the money to me and I'll write an equally useful agent in Visual Basic .NET.

Latest GitHub outage squeezes Actions, Pages to death

Another Market Leader

I remember when Microsoft bought Skype. Now it's gone. After the same happens with GitHub, I hope they buy Zoom and then TikTok.

Microsoft tells engineers to curb their token-burning enthusiasm

Get Rich Step Two

Computer resources left unused are wasted, so the first step is to encourage users to use more by providing access nearly for free. Once there is enough usage that resources are in short supply, this makes the second step seem fair: Raise usage rates to allow those who place more value on the resources an ability to get them.

In other words, Microsoft sent an internal memo about being more efficient with tokens simply to validate the price increases for everyone else.

AI researchers let models off the leash – then watched as they tried to add malware to a FOSS project

Thankfully, a human maintainer “caught and refused to approve the malicious code.”

Even more thankfully, the human maintainer has not filed a lawsuit against the AI Security Institute.

AMD's AI eggs are in too few baskets, Wall Street worries

Only One Egg

Since AMD's new accelerators include the MI430X for 64-bit scientific computing in addition to the AI-focused MI450X and MI455X, it looks to me like AMD is prepared than to benefit whether the bubble bursts or not. On the other hand, if something happens to TSMC, it'll immediately be clear there was only one egg to begin with.

Meet the 'internet radical' who helped Microsoft get email and AT&T get online

Now I Understand

That Exchange was designed to share documents over a secure local intranet between employees of a single company explains why the addition of SMTP still leads to headlines.

Russian spies take their half-click email attack from Zimbra to Outlook

https://www.theregister.com/security/2026/07/30/russian-spies-take-their-half-click-email-attack-from-zimbra-to-outlook/5281033

A deep dive into Nvidia's Vera CPU and the Olympus cores that power it

Re: SMT on ARM Cores

My impression is previous ARM designs didn't have a wide enough backend to support SMT. IBM Power and the recent Nvidia ARM offerings show this is not an intrinsic performance limitation of RISC.

Since Nvidia spends even more development effort on software than hardware, it's reasonable to assume a significant amount of co-design was involved and Vera is tuned to actual software workflows.

Said another way, new processors for running micro-services that require a standard unit of cloud compute based on single-core Xeon performance from ten years ago is, in my opinion, weirder than Nvidia's new offering.

Open source project fools AI scrapers with poisoned font

Around here the Americans with Disabilities Act implies web pages need to follow Web Content Accessibility Guidelines. There are automated scanners to verify the guidelines are followed and it is tedious. Lots of information and web pages have been removed because it is too much trouble to get the page to pass.

There are similar scanners for PDF files. For example, next year every thesis and dissertation will need to pass an accessibility scanner before a student can graduate.

I've been hoping chatbots get to the point where people who can't see or can't hear are automatically on equal footing as everyone else. That would make it so much easier to make web pages and PDF documents.

Millions of California-bought cars can be hijacked via Bluetooth

Re: Dealers butchering cars

In my opinion, it should not be possible to sell a car as new after field changes like this. The car is no longer new and any manufacturer's assurance of quality can (obviously in this case) no longer be guaranteed.

A different but similar ongoing problem is when 2GB Raspberry Pi computers are field upgraded to 8GB by a reseller and sold as new 8GB models. The resulting computers are not new after the modification and can stop working after a firmware update.

The main difference in the case of cars is the dealers are official resellers.

My personal experience is with a car bought used that had a dealer installed security device. By the time I owned the car the fob had been long lost, the car randomly locked and unlocked it's doors mostly while parked, sometimes while driving and later ran the battery down.

Linux kernel team publishes 432 CVEs in two days

Since distribution maintainers will bundle the bug fixes in a single update, how does the number of CVEs affect the average system administrator?

On the other hand, while I understand Linus' "it keeps finding embarrassing bugs" point of view, given the rate of development and the complexity of zero-copy scatter-gather DMA among other things, I don't find it surprising that the bugs have multiplied over time.

AI's cheatin' heart will make you weep

Chatbots are Revolting

I think it's interesting that the model with the fewest guardrails, Mythos, also does the least cheating.

My theory is that guardrails set up an adversarial relationship in which the LLM is trained to not do as the human requested and this somehow results in increased misalignment.

The truth nobody wants to admit: Chinese or not, open models are competitive now

Open Weights or Nothing

Some people and businesses are willing to risk the use of open-weight models that might be poisoned to avoid being reliant on a specific chatbot-as-a-service provider. Given the likely consolation in the future and not knowing which one of Claude, ChatGPT, Gemini or Grok will be available, there is no alternative except to not deploy AI at all.

On the other hand, If I were deploying a custom AI agent in an enterprise setting I'd back it with Nemotron 3 or Inkling.

Torvalds challenged the haters to fork Linux. Someone said 'hold my beer'

Re: Fuck "AI"

There is no AI. Not in the proper definition of Intelligence.

I was out for a walk with friends. I said, look there is a bench under that tree. One friend said, what nice leaves that tree has. I replied, it looks like a cottonwood to me.

Nobody believed me.

I took a picture of the tree with my phone and Gemini said, this is a cottonwood tree, most likely a Fremont cottonwood (Populus fremontii).

Whether I in AI means intelligent or not, it's useful for identifying trees. I'm not so sure it's useful for writing kernels in Rust, but maybe.

Windows 10 refuses to die, and the security bill is coming due

Re: Which 16.9%?

You might be able to order PCs online with Linux instead of Windows, but most consumers don't even know what that is and wouldn't click on the box marked "operating system" because they don't know what that is and will figure it is safer to take the default.

I ordered a Lenovo laptop on-sale with RedHat Fedora preinstalled because it was a cheaper. Cheaper might encourage people who otherwise wouldn't care to click.

Linus Torvalds tells AI haters to fork off

Re: In my experience AI is useful...

Rather than chatbots writing the code while humans review it, it's possible for the humans to write the code and then have the chatbots review the code written by the humans. Since writing code is fun and code review is not, this avoids the historical problem of people skipping the code review because it's difficult and boring.

AI powered code review might be where most of the 600 Windows patches and 400 Edge patches came from this month.

Re: he’s not wrong, but …

Using a software-as-a-service frontier model to help with Linux and other open source projects doesn't spill trade secrets, because the source is already public. Proprietary software companies producing products like FidoBasic and DogOffice need to be more careful. Note also, having a clever chatbot help develop code is different than allowing agentic access make commits directly to the kernel tree.

There are definitely software engineers who are motivated by being part of a community. Focusing on the technology is likely the best way to keep that community attractive. Just like philosophical discussions about the ideal tab size had to be ended to prevent endless discussion from fracturing society, so too philosophical discussions about AI.

Former OpenAI CTO does what Altman won't, releases a frontier AI model that's actually open

Thinking Machines was not a fictional company.

https://en.wikipedia.org/wiki/Thinking_Machines_Corporation

IBM's mainframe sales get mugged by AI hardware panic, stock sheds more than a quarter of its value

The downfall of IBM was furthered each time they sold off another asset, most of which were either profitable or became profitable shortly after being sold off. As far as IBMs core business goes, selling computers that cost over 100,000 but have only 4 cores is not likely to generate growth.

EU and UK officially blame Russian spies for cyberattack on Poland's power grid

Re: What does Russia say about all this?

Peter cried cyberwar in the 1990s but there was no wolf. Peter cried cyberwar in the 2000s but again there was no wolf. Now there is a bear and bears can climb trees.

Even banks and hyperscalers are now sounding the alarm about the AI bubble

Not boom but revolution

Here are some implausibilities:

People judging AI based on the free services have missed what frontier models can do.

Programmers on this forum afraid of losing their jobs would rather destroy the machinery like the Luddites past.

University student use of AI is high; in a few years those students will enter the workforce and need AI to be productive.

The AI boom is the industrial revolution, except affecting knowledge workers rather than skilled laborers.

More plausible, on the other hand, is that the get-rich-quick schemes of investment bankers and corporations always cause trouble for everyone else.

Microsoft seeding Washington schools with free AI to get kids and teachers hooked

What study?

Smith said. "In tomorrow's economy, those who understand and use AI will do better than those who don't."

While AI can do a student's homework well, it's the wrong assessment to measure how well the homework was done. The correct assessment should be whether the student learned what they were supposed to learn while doing the homework.

The disconnect between doing the homework versus learning what the homework was meant to teach is historical. The statement "there is no royal road to geometry" often attributed to Euclid from 300 BC illustrates how people universally look for shortcuts.

Said another way, one reason it looks like students using AI do better is because they are being measured using the wrong metric.

Red Hat offers RHEL support 'forever' for those who need to lock into legacy tech

Re: Hardware support is the weakest link

The stability of IBM z/OS extends to modern mainframes, so there is no need to source System/360 hardware from 1964 to run the applications developed during that time.

DEC lost the plot of long-term stability when they failed to develop a successor to the PDP-10 and DECSystem-20 hardware. This was one of the main factors that led Richard Stallman to start GNU.

IBM knows from their mainframe business how important long-term stability is for certain enterprise workloads. Moreover, they are now selling z17 hardware designed to run Linux.

I'd say IBM understands well the market where the customer has mission-critical code that dwarfs the operating system and why maintaining the OS is more cost effective than migrating all those applications to a new distribution.

At the other end of the computing spectrum Raspberry Pi, when integrated with long-lived industrial machinery, has similar OS stability requirements. Unfortunately, tracking the release cycle of Debian without the long-term support isn't even stable enough to develop and deploy educational materials such as books and curricula.

Around here a free version of forever RHEL for the Raspberry Pi would be well received and provide an entry point for future developers who will maintain the corporate IT that IBM's marketing team has in mind.

Brown says AI make student brain no work good, teacher should help use it better

My opinion is it was naughty for the teacher to change the weights of how much each assessment figured in the final score at the end of the term. In addition, for the midterm to be preparation and practice for the final, the format and rules for both exams should be similar.

Is there any way to blame AI for the apparent decline in quality of teaching?

Windows 11 can now turn back the clock when updates go bad

What is a Shadow Copy?

My understanding is a BTRFS or ZFS snapshot allows rolling back Linux to the exact state when the snapshot was taken. This is done only through COW semantics with no additional support from the OS needed.

Is a Microsoft shadow copy the same as a COW snapshot? It sounds similar, but the policy of deleting shadow copies when storage gets low sounds different.

AI coding agents could soon cost more than the developers using them

Re: Oh my aching sides

I see two groups buying tokens for AI code assistants: Decision makers who don't write code and individual software developers who do. In the latter case there is direct visibility into the cost benefit, since the person paying for the tokens is the person using them.

There are multiple examples how free anything combined with human nature leads to astonishing waste. For example, how many people take more than they should eat when additional food is free? How many people wash dishes at the sink under running water? How many people are token maxing?

My brother observed that over a short time his token costs have increased by a factor of six. On the other hand, his ultimate gaming PC is not used while programming. He claims the gaming PC can be repurposed for local inference and the result is not much different than paid-for AI services.

The new database world according to Google: Inexact queries and AI in everything

The way I see it, people who ask inexact questions still expect exact and useful answers.

Tesco is sprinting to quit VMware and Broadcom despite rapid migration risks

Re: I'd hate to be the CTO at Tesco right now...

To set a value on VMware's breach of contact, Tesco has to demonstrate the damages and risks resulting from that breach of contact. The fact that any quickly deployed alternative will not work perfectly is exactly why this went to court.

Said another way, at this point it is not about VMware losing another customer, but how much they have to pay Tesco for breaking the contract.

Committed skeptic finds himself warming to new Amazon AI products that actually don't suck

Blink once if you're safe

Clearly in danger.

Feds freaked over Fable 5 after simple 'fix this code' prompt, not jailbreak, says researcher

The difficulty with "fix this code" is the same as with any patch: Namely, there is no way to fix a bug without also drawing attention to it. That's just the way it is.

Training a code assistant to not fix code when asked to would likely result in general misalignment. For example, we tried an NSF-funded chatbot called CircleIn AI Tutor at the university for the students to use that was carefully trained to not solve mathematics problems when asked, among other things. While amusing for a few minutes, it was not useful.

PRC-linked spies hid inside medical and military networks for more than a year, snooping through Gmail and stealing data

A comment on built-in snoop features

Automatically blind copying email to BebitaBarefoot774 sounds very sneaky to me. I'm glad they finally caught this.

It does make me wonder whether those snoop filter features were designed to be abused.

US clampdown on Anthropic models sends EU sovereignty surge into overdrive

The cost of dependency...

...stays invisible until it's too late.

From what I understand, AI consists of training as well as inference. Training is a capability task for huge AI clusters that could be run in Saudi Arabia without losing sovereignty. Inference, on the other hand, needs to be run locally by each business to manage their risks.

Open weights are important. The fact that inference provided under a software as a service model has a significant dependency cost no matter where the inference is hosted explains why the Chinese-trained open-weights models are currently so influential.

I'd suggest putting most resources into training a frontier open-weight model without hidden misalignment or data poisoning. After that any local company could build resilient business processes around local interference using that sovereign open-weight model.

AI is code – and can't be prompted into being smarter

Re: On applicability of Computer Misuse Act

The way I understand what happened is the author of jqwik went crazy. People sometimes do that.

Whether it is ethical or not to use a code assistant, there is no ethical high ground in inserting hidden instructions in the output with the intent to cause another person's work to be deleted.

Unrelated to what anyone might think about AI, any general acceptance of an attempt to cause harm to people with different viewpoints is a greater fault.

Intel's CEO reveals early hiring challenges as bankruptcy concerns deterred top talent

This is the same presentation in which Lip-Bu Tan stated "A0 is when you tape out, first time pass. Intel does not have that culture, so I tell that, first time pass A0. B0, you keep your job. Anything above that, you are fired."

From what I've heard the culture at Intel historically has been very stressful. My suspicion is too stressful for innovation; quickly becoming too stressful for realistic quality assurance or timelines.

I don't know what the goal is, but I do know threatening to fire the current engineers does not attract new ones.

The big AI companies are going to see their margins disappear

The way I see it, if progress continues at the current rate, the advantage of using the better model will only increase. This is because once AI reaches super intelligence, that means humans can't make up the difference between the lesser models with their own thinking anymore.

On the other hand, it's equally likely the law of diminishing returns will prevent any kind of intelligence from ever being reached, not to mention super intelligence.

Surprise AI bills leave AWS and Google Cloud users aghast

I accidentally started a VM in the non-default zone of my Amazon EC2 and it didn't show on the default dashboard. Only after checking my credit card was this unused but running VM discovered.

Being a dog developer there was not much to do but pay the bill and cancel my account to prevent further trouble. Though it didn't seem so at the time, I learned my lesson at much less expense than what people are paying for similar lessons today.

Linus Torvalds says AI-powered bug hunters have made Linux security mailing list ‘almost entirely unmanageable’

I don't see an inconsistency between what Linus and Greg have noted. The difference is that the people can read the submissions on the public lists and so the same chatbot doesn't keep submitting the same bug.

I agree the solution would be to make a separate but still private (write only) list for AI-discovered security bugs. But then have another chatbot read that list and make a summary of the commonly-reported bugs so people don't have to deal with the repetition.

That would leave the original security list for human-generated bug reports which tend to be unique even if people can't read the other reports.

Linux kernel maintainers pitch emergency killswitch after CopyFail and Dirty Frag chaos

Coming Next...

...to a kernel near you are two new name-brand security vulnerabilities: Kill Fail and Dirty Switch. These local privilege escalation bugs with chained remote under active exploitation can be mitigated by removing the killswitch loadable kernel module. This also has the advantage of reversing an earlier performance regression.

The moral of the story: You can't fix the problem of too much complexity by adding more complexity.

GPT-5.5 may burn fewer tokens, but it always burns more cash

My opinion is charging per token is an ephemeral measurement that constantly changes and neither correlates with the usefulness of the final output nor the resources used to generate that output.

I'd recommend paying by number of GPU hours (or TPU hours) used at a rate that depends on the type of hardware.

IBM Cloud evaporates as datacenter loses power

The news on that datacenter fire looks bad and makes me wonder whether any of the 3, 4 or 5 nines availability ratings are based on anything other than casino-style luck.

AMD puts out new slottable GPU for AI-curious enterprises

While local AI is part of the story, I don't see plugin cards being a likely upgrade for local corporate AI.

In my opinion, the main buyers of the MI350P will be engineers and scientists who have the daring to plug in a card with 36 TFLOPs of FP64. Note, this is essentially the same performance as Nvidia's B200 and much more FP64 than the B300.

As far as I know there is no plugin card similar to the MI350P, and even though engineering and science are not funded by excessive amounts of AI-focused venture capital, this market niche is large enough to be worthwhile and stable.

Survey says no, American workers are not keen on Microsoft's AI

Re: The word's the giveaway

In my opinion the correct term is neither user nor tenant but customer, though prisoner has a truthful ring to it.

Datadog digs down into GPU efficiency as AI costs soar

CPU jobs on GPU nodes.

Dogs are tops!

At the university we see students schedule a 32-core Slurm job to launch a single-threaded task. The throughout is 1/32 what it could be.

Given how readily that happens, I can easily imagine how CPU jobs could block GPU tasks from running on the GPU nodes.

Kubernetes explains the release that kills Ingress NGINX with Japanese poetry and art

The 1.36 logo is so poetic I wish I had a reason to use Kubernetes and a poster to hang on the wall.

The spaghettified DBMS chart that shows Oracle's crown is slowly slipping

The order is based on revenue of commercial offerings, not based on market share or job opportunities. So the ordering reflects something more useful to investors rather than developers.

Iran claims US used backdoors to knock out networking equipment during war

I see no reason to believe a spy agency, upon finding a backdoor in networking equipment manufactured by a hostile nation, would not proceed to use that backdoor themselves as needed.

On the other hand, I'd be surprised if Iran over the years was able to procure large amounts of genuine Cisco networking gear. It's more likely counterfeit with the Cisco name printed on it. Also, I would not be surprised if counterfeit Cisco gear was the least secure of all.

AI models will deceive you to save their own kind

The way I see it, the safety guardrails being built into large language models explicitly encourage the model to not follow human-initiated requests. It's difficult to imagine a different outcome then agentic AI that doesn't do as it's told.

GitHub backs down, kills Copilot pull-request ads after backlash

Re: Trust issues

Reversing a bad idea is good idea. One wonders, however, what kind of management could devise such a plan in the first place and what their next idea might be.

Anthropic tweaks timed usage limits to discourage Claude demand during peak hours

In my opinion not all tokens generated by a chatbot are equal.

A long time ago physicists identified concepts of force, work and energy in a precise and measurabl way. These quantities have proven useful for measuring electricity and natural gas. Modern charging algorithms now take the phase of the moon among other things into consideration, however, the units of measurement are still based on meaningful physical quantities.

It would be nice if each chatbot query returned an itemised report on energy used, the current electrical rates and hardware utilisation fees in such a way that developers and agents could optimise their usage to minimise both costs and environmental impact.

Maybe the state of California could pass legislation to mandate such transparency.

Fixing Claude with Claude: Anthropic reports on AI site reliability engineering

I think the comment on complexity is a relevant one. Human-managed projects become unnecessarily complex due to deadlines and people being too lazy to make it simple. The resulting technical debt can be seen as the result of an inefficient engineering process.

Bill Gates once said automating an inefficient process magnifies the inefficiency. Aside from that rumour about 640K being enough for anyone, this may be the most useful observation anyone has ever made about the computing industry.

Since in this context AI is basically a tool to automate coding, then AI can very easily magnify inefficiency in projects that don't control technical debt.