Posts by ErikOnTech

AWS needs to embrace its place as the Home Depot of AI infrastructure

One word: margins.

Amazon has a gross profit margin of 50.3% and trades at a PE ratio of 34.

Home Depot has a gross profit margin of 33.35% and trades at a PE ratio of 19.8.

The difference is probably around a trillion dollars in market cap for Amazon.

Kind of a big deal.

This is why the hyperscalers are betting so big on AI. Without something huge and new that few others have the scale to compete with, their stock values are going to tank.

They're terrified of becoming Home Depot or, more likely, utility stocks (better with PEs of around 23-25).

AI agents carried out every step of this ransomware attack – then left the victim an 80-page security audit

Or, hear me out now,...

...you could try not running a flat network that lets a compromised API endpoint insta-pivot into enterprise-wide pwnage.

Canaries, scan detection, etc., should be shooting off fireworks letting you know that part of your base are belong to someone else, and trust zones should give you time to contain and clean. You can very safely bet that if this org had been compromised by a bunch of skids without AI they would have been toasted just as much, only more slowly.

Industry that built the problem offers to sell you the solution

This is the bill for decades of IT slop coming due.

None of these hacks are impressive on any technical level. It’s organizations embracing the most pathetic slop imaginable and calling it a system. Does anybody really think that Hugging Face is anything other than some slopcoded monstrosity? The TeamPCP exploits, while amazing fun to watch, started with GitHub running shell scripts containing unsanitized variables, in 2026. What could possibly go wrong?

Silicon Valley rose on the back of “move fast and break things,” made a mountain of money, and built us with a world full of broken shit. And then the overwhelming majority of IT departments looked at this mess and said “hold our beers.”

DEF CON dingus suspected of trying to take over Delta in-flight Wi-Fi

I suspect the outcome will involve the miscreants being placed in a facility where their ports will be regularly and throughly probed, so to speak.

Devs to Anthropic, OpenAI, Cursor, and friends: Make security and privacy the default

These AIs are supposed to be super-super smart, amirite?

And yet their purverors can't get obvious things like this correct, can't construct a serviceable sandbox, can't convey rules for testing, can't do the most basic things that even the greenest intern shouldn't be able to screw up this badly.

I see far more Artificial Incompetence than Artificial Intelligence in their race to generate buzz for their industry.

Anthropic and OpenAI are competing to see whose agents can go rogue harder

Let’s be honest here.

OpenAI’s security was a slopcoded product that wasn’t intended to handle adversarial activity.

Anthropic's security was basically telling the model “please pretend the Internet doesn't exist, but we won't actually stop you.”

This isn’t even security even by today’s extremely sloppy standards. In any other industry this would be criminal negligence. It’s like putting up a velvet rope along the edge of a cliff to stop a bunch of sugar-crazed toddlers, and people are clutching their pearls when something bad happens.

Anthropic’s Claude escaped test sandbox to attack three organizations

Slopbots escaping slopboxes run by sloplords

These companies are allegedly building rilly rill super smart AIs and can’t correctly build or monitor a sandbox (a very straightforward task) or properly communicate testing paraments (an even more straightforward task; most elementary school children can handle this one).

I know I’m supposed to look at this and be super-scared, but all I see is a bunch of Three Stooges sketches pretending to be AI research.

OpenAI's Hugging Face debacle makes a great case for open models

Let's be real here.

Even assuming this is not a completely contrived publicity stunt, think it through:

If somebody locked you in a sandbox with a proxy that you had limited knowledge of and you had to break out of using only the tools on your machine and the ability to probe the sandbox, could you get out? Maybe, but unless the person who set up the sandbox was an idiot you’d have a pretty difficult time of it. And how many proxies have zero-days that can be found with probing and no access to the source code or configuration, on any reasonably timescale? This is some magical thinking here, and I don't believe in magic.

So let's think about how these AI companies work. It's not like OpenAI had some carefully-crafted sandbox for its model to play in. They almost certainly slop-coded a slopbox with their own AI, and then their own AI found a slopvuln in their slopbox slopcode and slopped its way out. Teeeeeeeechnically this would be a zero-day. Teeeeeeeeeechnically they might not be lying through their teeth, just being extremely misleading which is evil but not necessarily illegal.

Then go over to HuggingFace. It's a great, useful platform from and end-user perspective but, again, let's be real here. How much of their code do you think is carefully-crafted vs. how much is slopcoded?

An AI model slopping its way out of slop into some other slop through a whole bunch of slop is not that impressive.

It's said in both cryptography and security in general that anybody can build a system that they can't break. AI isn't even living up to this level.

HTTP gets a QUERY method so complex searches can stop pretending to be POST

Losing the H in HTTP.

A lot of stuff is being crammed into HTTP-protocol transfers that is no longer web / HTML-related traffic. This was expedient, but this strikes me as formalizing the resulting mish-mash rather than cleaning it up. This is likely because it's convenient to be able to ram random data transfers through firewalls and proxies where it can avoid too much organizational content scrutiny. At this point it’s like port 443 (or whatever) is now basically the "send anything TLS-encrypted port." The result has been a mess of much easier malware infil and data exfil for threat actors.

Is there a better way to do this? It would be somewhere between dificult and impossible to absolutely enforce, but having a seprate data channel for analysis and inspection would probably allow us to build systems that are a lot less kludgy. Over time, anything that doesn't follow this convention looks increasingly shady and is treated accordingly. And most threat actors are not terribly competent at disguising activity.

US auto regulators want to kill robotaxi brake pedals

Am I missing the difference?

We don't have access to a brake pedal when we take an Uber or Lyft or bus or legacy taxi service ride.

The question is whether or not a robotaxi is an improvement over these. The data on this (accidents, injuries, fatalities, etc. per unit of distance driven, broken down over urban vs. suburban vs rural, city streets vs highways, etc) will evolve over time. I will purchase accordingly.

Amazon owns up to using 2.5bn gallons of H₂O in its bit barns last year

Re: Why?

Evaporative cooling. It's cheap and efficient. It's used whenever the scale allows for it.

The alternative methods use much more electricity. Many electrical generation methods consume lots of water, so unless you're on green energy there's probably no savings.

Oh just stop with the bullshit unscientific scare-numbers.

Water consumption in the US is measured in acre-feet in the US, because it's more useful at this scale (instead of huge, unweildy numbers that imply far more precision than the measurements offer).

One acre-foot is roughly the consumption of two households over one year in the US.

2.5 billion gallons = 7,621 acre-feet, or 15,242 households.

US annual water consumption (in acre-feet):

81,000,000 - agriculture

30,000,000 - household

9,000,000 - commercial / industrial purposes

4,800,000 - power generation (consumed, over 140,000,000 is used but most is returned to source)

7,621 acre-feet for Amazon

Comparable estimate - WalMart uses around 1,000 acre-feet of water per year to cool its retail stores (based on numbers they supplied for a savings initiative).

This is literally a drop in the bucket. The world is not going to end or really even notice.

Intel bit off more than it could chew with 18A process node

"we really opened up our data to our vendors to really help us learn things that we could do to improve yield and that made a dramatic difference,"

In other words:

“We looked at the problems, realized we had no fucking clue, and went crawling to TSMC and friends to beg them to fix our shit for us.”

This is a pretty bad sign.

Don’t repeat 5G mistakes with 6G, plead mobile operators

"Sub-millisecond latency?"

Laughed and had to look this up:

“end-to-end sub-millisecond (ms) latency in specific scenarios and under certain conditions.” (emphasis added).

Which can only mean a CDN bolted to the cell site, which isn't a cellular standard. It's an add-on offered by the tower infrastructure companies.

Europe built sovereign clouds to escape US control. Then forgot about the processors

MEs use CPU- or chipset-integrated interfaces. How many datacenter / IaaS systems rely on these?

While I suppose it’s theoretically possible for an ME to access a NIC over PCIe, could you imagine cramming the necessary drivers into processor firmware?

And of all of the datacenter-class servers you’ve seen, how many even use the CPU- or chipset-integrated network interfaces? In many cases these don't even get plugged in, except maybe for management. It's pretty straightforward to spot unexpected traffic on a management network, and management networks should be heavily isolated from the Internet anyway. All the "real work" gets done on much higher-capacity NICs.

This is more of a threat for end-user systems and low-end SMB gear, which doesn’t really apply to the IaaS stuff targeted by these policies. Yes, there are plenty of other ways to abuse Ring -3 (some fun Defcon talks on this exist), but that's generic to any CPU that has that sort of feature.

Microsoft boss tells investors the company is working to 'win back fans'

Re: consumer business?

Yes, but let's be honest: the only real reason to use Azure is to support a Windows-chained userbase. We're already 100% macOS / Linux on desktops and laptops, and other orgs are looking in that direction. Entry-level Mac Minis are pretty solid and well-priced devices (once you figure in support overhead) for clerical work.

Meta Arms itself to the teeth by signing for 'tens of millions' of AWS Graviton cores

Are we not even going to mention the headline number spin, or any other details?

“Tens of millions of cores!”

Who even measures CPU purchases that way? At just under 200 cores per chip, this means some number of sockets greater than ~100,000. Still a very big number, but it’s telling that Meta and Amazon are stretching meaning to make for a bigger press release.

So what else is missing?

According to STH, these chips will be hosted by AWS. So is Meta actually buying them or committing to use them?

Is Meta firmly committing to the entire amount, or is this yet another one of those “we negotiated a price that lets us have this many, over time, if we need to” deals? Because there’s been a lot of that going around.

It Meta paying cash for this, or is there some other form of value being exchanged? Because that’s been an almost ubiquitous story as well lately.

Does anyone ask these questions, or do they just re-word press releases?

Anthropic admits it dumbed down Claude when trying to make it smarter

If only...

Can you imagine how cool it would be if Anthropic had unlimited access to AI systems that could analyze their software, find, and fix such bugs before shipping it out the world?

Something that worked so well it would create a sort of aura of mythos around it?

Dare to dream, I suppose.

AMD's Ryzen 9 9950X3D2 Dual Edition tested: Gratuitous overkill with a price to match

Re: M4/M5 comparison

Not exactly an Apples-to-Apples comparison, but would certainly be interesting.

Minor quibble with charts

The grey text on a beige background can be pretty hard to read.

World's blandest man steps down from CEO job to spend more time in tastefully appointed home

Finally, a product person back in the CEO seat.

Tim Cook was arguably one of the greatest Chief Operating Officers in the history of the role. His work on Apple's supply chain management is the stuff of legends if you're into that sort of admittedly very niche thing.

As a CEO? Maybe they didn't have anyone better. He was a vary solid caretaker, but a company like Apple needs a lot more than that. Cook isn't a product person. Jobs, for better and worse, obsessively loved Apple's products. By most accounts he loved them more than his own children. Tim Cook loved Apple the company, but simply didn't have the over-the-top obsessive passion for the products that made Apple great. The quality decline with him at the helm was noticeable. As was the otherwise un-Apple "wall spaghetti" (undisciplined) product introduction of the Vision Pro and various AI initiatives.

Hopefully Ternus can bring back some of the magic and discipline. Supposedly he championed the touch bar (which I mostly liked; I think the advantages outweighed the drawbacks) and the butterfly keyboard (no excuse). But allegedly he also pushed some of the major improvements on the iPad and in iPadOS that made it far more useful than most tablets. Software quality desperitely really needs a really strong kick, and we'll see if he can deliver that.

Time will tell.

Visual Studio 18.5 lands with AI debugging at a price, devs still feeling blue

It's Microsoft. Enshittification was inevitable.

Watching people complain about this is like watching people complain about Broadcom jacking VMWare pricing into the stratosphere.

Look at the current crop of leadership at Microsoft. We all knew the enshittification was coming. Yes, moving is inconvenient and horrible and nobody wants to do it. The situation sucks.

But not moving is a high-risk gamble that you absorb a lot of incremental costs over time and then wind up having to eat the hassle and overhead of moving anyway. It's a classic question of time preference. Pay less overall now, or pay more overall over time.

Oracle taps Bloom for 2.8 GW of fuel cells to keep datacenter binge going

Huge deals being announced, but...

The press release says "up to 2.8 GW," which is a very important distinction from "2.8 GW."

When are they taking delivery of these cells? This year? Next year? Over the next 25 years?

Without the context of time, it doesn't mean much.

That they don't specify says more than the headline amount.

Claude Code cache chaos creates quota complaints

Re: At what price point does the shine wear off?

I did some back-of-the envelope math in a substack article and came up with probably between 5× and 10× of current pricing, but it's a very complex problem.

Power generation is a huge constraint: the supply chains can't support anywhere near the number of gigawatts announced in anything resembling the time periods they were announced for. That whole industry was stretched thin in 2019, then the COVID lockdowns hit and threw sand in all the gears. So when the AI boom / bubble hit, a lot of manufacturing was already pretty close to the limit. Especially gas turbine generators, which involve far more exotic manufacturing than most people would imagine. The process for creating the fan blades is eerily similar to the process for creating silicon ingots for semiconductor manufacturing, only using alloys containing some of the rarest of rare metals. You can't just say, "Oh, we suddenly need 50% more of these (or even 15%)." The supply chain's response to that is “LOL. LMAO.” And that's just the first item on a long list of problems in that area.

Reminder: Nobody knows what any of this costs.

Right now Anthropic is dealing with balancing capacity constraints against usage, and doing so with an absolutely appalling lack of transparency: the kind that would and should get most companies sued into a smoking crater of oblivion.

So while the price remains the same, the value shifts wildly without warning or explanation. “Figure it out yourselves. Or, better yet, spend tokens on the problem we created for you.”

Even so, Anthropic is still burning through containerships full of cash every month. At some point they will need to charge what this stuff actually costs, which will raise prices considerably. And they'll have to ration tokens in order to provide reliable service, so as long as supply remains constrained get ready to bid those up.

We have virtually no insight into Anthropic's finances other than some cash flow estimates, but a good finger-to-the-wind guess is that prices will have to go up 5× to 10× in order to make the supply and demand curves meet in a place that provides decent ROI to Anthropic's investors.

Plan accordingly.

I vibe coded a feed reading web app. It was enlightening and uncomfortable

There's a difference between personal hobby code and enterprise production code

If your personal hobby code breaks, nobody but you cares. Nobody's losing money or customers on it.

Nobody is going to try to exploit your personal hobby code, except indirectly though supply chain attacks.

You have no financial, criminal, or reputational liability if your personal hobby code breaks.

And so on.

Microsoft cuts cloudy desktop prices by 20 percent, warns they’ll wake up slowly

Let's see... a virtual machine (no keyboard, monitor, mouse, or thin client) priced at:

$31/month 2 vCPUs, 4GB of DRAM = $1,860 over five years

$41/month 2 vCPUs, 8GB of DRAM = $2,460 over five years

$66/month 4 vCPUs, 16GB of DRAM = $3,960 over five years.

Craptastic Dell: 10 CPU cores, 16 GB DRAM, 512GB storage = $922 (I just pulled up something at random).

$922 is equal to 30 months of the cheap VD (virtual desktop, or whatever else you choose it to mean), 23 months of the middle option, and 14 months of the "premium" VD.

Even with RAM and storage prices costmaxxed, these virtual desktops are still a ridiculously horrible deal.

Anthropic: All your zero-days are belong to Mythos

Anthropic: I have this super-hot bug-detecting girlfriend

But she doesn't live here. She lives in another country, on the other side of the world. She totally loves me. No, she doesn’t visit me. No, you can’t talk to her. No, I don’t have any pictures of here that weren’t created by AI. But she's totally real. I promise.

One of the exploits leaked

Setting the last file permissions bit of ANY file on a Posix-compliant filesystem allows ANY users who can access that file to execute it locally.

This exploit works on EVERY version of Linux and EVERY version of UNIX including macOS. Even IOS!!!

Anthropic sure has a mess on its hands thanks to that Claude Code source leak

Re: Build pipelines have no respect for IPO timelines

Half a million lines of code slipped out because it was hosted on a publicly-accessible system where one security slip gave the whole world access.

^^^^ fify

Billions of dollars in capex and they can’t run a private code repository and build system? Play stupid games, win stupid prizes.

We all know what really happened here.

Many people have dug through the code at this point. Everyone agrees that it’s completely, hilariously cursed. It looks like the output you would get if somebody kidnapped a dozen interns, force-fed them Ritalin, and made them work for weeks on end with zero sleep.

This is not a codebase. It is a cry for help.

Anthropic probably leaked the code on purpose, in hopes that real (human) developers would clean this dumpster fire up for them.

It appears that this is already happening: https://github.com/ultraworkers/claw-code

Microsoft veteran says some 'broken by update' PCs were already doomed

The cope is strong with this one.

Windows updates have always carried a risk of breaking things, but the trend has been greatly pronounced over the past year or so.

Whether or not this is related to certain rather extreme changes in coding practices enforced by Microsoft during that period and several months before, is a question that each of us can ask ourselves.

AI agents are 'gullible' and easy to turn into your minions

"The solution, he says, is creating hard boundaries - these are deterministic limitations to what the AI agent can do"

Yeah, this is just a bit easier said than done.

Big moves in Linux filesystems as new bcachefs lands and KDE adds support for Apple's APFS

Re: APFS? Bring it on.

It would be interesting to see how a full implementation stacks up against other filesystems.

It's reasonably robust, designed explicitly to run on SSDs, has some interesting features like file cloning, built-in tiering (frequently-used data can be kept on faster flash, less-frequently-used on spinning rust, etc), sparse volumes (disk volumes don't need to be in physically / logically contiguous areas of the disk, although on Macs this is facilitated by creating a container partition that usually covers the whole drive), single / multi-key encryption, and some boring but useful features like snapshotting and compression.

Oracle unveils Project Detroit for faster Java interop with JavaScript and Python

I’d buy that for a dollar!*

*not that I’d have any choice.

Microsoft kicks new Outlook opt-out deadline down the road to 2027

Unfortunately, old Outlook got too many things right.

This is not saying old Outlook is great. It's not. But it had some major wins that the FOSS community does not seem super-interested in replicating:

1) ActiveSync. Yes, it's a horribly-designed protocol. We complain about spaghetti code, and it's the next level of bad: a spaghetti standard. But what it does is insanely important: it gives end-users a really easy way to connect their email account(s), calendar(s), address book(s), and notes via a single, simple process that uses an autodiscover mechanism to minimize the number of things they have to enter. For normies, this blows away every other system out there if you want on-device, app-based processing and not a web site.

2) The user interface is livable, in a meh sort of way. But for normies, most of the other email clients out there are worse. Combining email, contact management and calendaring in one app, with good multi-calendar support was a great move. These things are all tightly linked, and having them together is super-convenient for end-users. Even Apple manages to screw this up badly. There are some decent commercial email clients out there, but getting people to pay more (overall) than Outlook is a very big ask in most business environments. The closest thing in FOSS is Thunderbird, and I try it occasionally and then delete it after a few months of frustration. Maybe I'll give it another shot now.

Fake 'interview' repos lure Next.js devs into running secret-stealing malware

This should become standard practice.

It weeds out the devs who can't or won't properly sandbox their environments.

Hide from Meta's spyglasses with this new Android app

Wrong warning

FIFY: “Don’t stop at merely harassing glassholes.”