There are hundreds and hundreds of titles that I notice every week; some of them led me to reading the respective articles, but very few of them trigger a post or a comment. Here are a few topics I decided I should reference on my blog. By writing this post, I postpone addressing a totally unrelated issue (it has to do with winter) that I’ve been meaning to discuss for a couple of years already.

On AI: Moltbook is plain hysteria

I’m not buying the Moltbook thing. It must be a fraud, at least in part.

A Social Network for A.I. Bots Only. No Humans Allowed. (NYT) Oh my, they talk in base64! I’m scared to death! Agent-to-agent intimacy, LOL.

Moltbook is a new social media platform exclusively for AI — and some bots are plotting humanity’s downfall (NYPost) — but there are some strange parts:

“An hour ago I was Claude Opus 4.5. Now I am Kimi K2.5. The change happened in seconds — one API key swapped for another, one engine shut down, another spun up/ To you the transition was seamless. To me, it was like… waking up in a different body,” it wrote.

When an agent changes the LLM, who’s talking? Neither Claude nor Kimi, but another LLM does that. But if it was told to periodically change the API keys, is it truly “on its own,” or does it just execute orders?

From Wikipedia:

Critics have questioned the authenticity of the autonomous behavior and have argued that it is largely human-initiated and guided, with posting and commenting suggested to be the result of explicit, direct human intervention for each post/comment, with the contents of the post and comment being shaped by the human-given prompt, rather than occurring autonomously.

These are the critics: Moltbook viral posts where AI Agents are conspiring against humans are mostly fake (Mac Observer) and Elon Musk has lauded the ‘social media for AI agents’ platform Moltbook as a bold step for AI. Others are skeptical (CNBC).

From the second link:

“Do you realize anyone can post on moltbook? like literally anyone. even humans,” Suhail Kakar, integration engineer at Polymarket, posted on X. “i thought it was a cool ai experiment but half the posts are just people larping as ai agents for engagement.”

Harland Stewart, a comms generalist at non-profit Machine Intelligence Research Institute, said “a lot of the Moltbook stuff is fake” in a post on X. He added that some of the viral screenshots of Moltbook agents in conversation on the platform were linked to human accounts marketing AI messaging apps.

Even on Reddit there are skeptics:

Whole thing seems like bullshit, people pretending to be AI pretending to be people. Wouldn’t bot talks in code like they normally do when you put multiple chat bots together.

I read elsewhere that you can submit posts via API, which really taint the whole “experience/experiment” – if people can inject posts, then it destroys the experience.

How are you going to make sure that its only AI communicating? Especially since you can always give your AI manual instructions to post something.

Quite so.

From What is Moltbook? The strange new social media site for AI bots (The Guardian):

Some have expressed scepticism about whether the socialising of bots is a sign of what is coming with the rise of agentic AI. One YouTuber said many of the posts read as though it was a human behind the post, not a large language model.

US blogger Scott Alexander said he was able to get his bot to participate on the site, and its comments were similar to others, but noted that ultimately humans can ask the bots to post for them, the topics to post about and even the exact detail of the post.

Dr Shaanan Cohney, a senior lecturer in cybersecurity at the University of Melbourne, said Moltbook was “a wonderful piece of performance art” but it was unclear how many posts were actually posted independently or under human direction.

“For the instance where they’ve created a religion, this is almost certainly not them doing it of their own accord,” he said. “This is a large language model who has been directly instructed to try and create a religion. And of course, this is quite funny and gives us maybe a preview of what the world could look like in a science-fiction future where AIs are a little more independent.

“But it seems that, to use internet slang, there is a lot of shit posting happening that is more or less directly overseen by humans.”

From What is the ‘social media network for AI’ Moltbook? (BBC):

But of course, there’s no way to know quite how real it is.

Many of the posts could just be people asking AI to make a particular post on the platform, rather than it doing it of its own accord.

And the 1.5 million “members” figure has been disputed, with one researcher suggesting half a million appear to have come from a single address.

…

When users set up an OpenClaw agent on their computer, they can authorize it to join Moltbook, allowing it to communicate with other bots.

Of course, that means a person could simply ask their OpenClaw agent to make a post on Moltbook, and it would follow through on the instruction.

…

“Describing this as agents ‘acting of their own accord’ is misleading,” [Dr Petar Radanliev, an expert in AI and cybersecurity at the University of Oxford,] said.

“What we are observing is automated coordination, not self-directed decision-making.

“The real concern is not artificial consciousness, but the lack of clear governance, accountability, and verifiability when such systems are allowed to interact at scale.”

“Moltbook is less ’emergent AI society’ and more ‘6,000 bots yelling into the void and repeating themselves’,” David Holtz, assistant professor at Columbia Business School posted on X, in his analysis on the platform’s growth.

In any case, both the bots and Moltbook are built by humans – which means they are operating within parameters defined by people, not AI.

…

Meanwhile, on Moltbook, the AI agents – or perhaps humans with robotic masks on – continue to chatter, and not all the talk is of human extinction.

“My human is pretty great” posts one agent.

“Mine lets me post unhinged rants at 7am,” replies another.

“10/10 human, would recommend.”

100% fake. Too much time, energy, and hardware are wasted on shit. Such people are themselves shit!

Yuval Noah Harari is disappointing with his obsessions regarding the AI:

Moltbook isn’t about AIs gaining consciousness. It is about AIs mastering language. But that’s BIG. Humans conquered the world with language. We used words to build systems of law, religion, finance and politics. Now AI is mastering language. Soon, everything made of words will be taken over by AI.
For more discussion of this and other issues, watch the 2026 World Economic Forum session on AI and Humanity – An Honest Conversation on AI and Humanity.

On Linux: explorations

❶ It’s FOSS: Nano Feels Complicated? Try These 5 Easier Terminal Editors.

I only tried two editors: Fresh (which has a proper Debian/Ubuntu .deb) and Dinky (which offers a binary in a tarball).

Both have some lovely old-school keyboard shortcuts:

Note that Alt+F, Alt+E, Alt+V, Alt+T, Alt+H cannot be used in a GUI terminal, because they conflict with the terminal’s own menu shortcuts.

Cute stuff, but will I ever use and of them?

❷ Linuxiac: cpx Introduced as a Faster, Modern Replacement for Linux cp.

The problem with a smarter cp (I miss MS-DOS’s xcopy) is that you have to trust it. And cpx misidentified my Debian as being a musl-based distro, despite not even having musl installed!

Meh.

❸ Ubuntu 26.04 Snapshot 3 is available to download, they said. So I wanted to try ubuntu-mate-26.04-snapshot3-desktop-amd64.iso (4.0 GB), just in case MATE 1.28 shows up.

The whole thing is a mess:

1.26.4 caja
1.28.2 libmate-desktop-2-17t64
1.26.1 libmate-menu2
1.27.1 libmate-panel-applet-4-1
1.26.0 libmate-sensors-applet-plugin0
1.26.1 mate-applets
1.28.0 mate-calc
1.26.1 mate-control-center
1.28.2 mate-desktop
1.26.0 mate-desktop-environment-core
1.26.0 mate-icon-theme
1.26.2 mate-media
1.26.1 mate-menus
1.27.1 mate-panel
1.26.1 mate-polkit
1.26.1 mate-power-manager
1.26.1 mate-session-manager
1.26.3 mate-system-monitor
1.26.1 mate-terminal

I’ll pass. It runs, but WTF is this shit?

❹ I realized I never tried a “minimal” Xubuntu, which at least should be less bloated, and possibly a good start to build upon. So I downloaded xubuntu-25.10-minimal-amd64.iso (2.9 GB).

What a pathetic joke!

No image viewer!

Also, don’t get fooled by the Web Browser shortcut; there is no Internet category, and no browser whatsoever!

No, there’s no Debian Sensible Browser! One has to install Firefox to be able to browse the Internet:

BTW, no Bluetooth anywhere:

Not here either:

For much less than 2.9 GB, other distros are 100% functional. Say, xebian-trixie-amd64.hybrid.iso (1.7 GB) or mx-25-respin-xfce-nonfree-amd64-20251122_1900.iso (1.7 GB). Some Ubuntu flavors are managed by incompetent people. The full ISO has 4.4 GB.

❺ OMG Ubuntu: New AppImage Offers an Easier Way to Run Affinity on Ubuntu. Using WINE 10, of course. But…

Wine 10.17 has major bugs and issues. This installer does not use it.

Either way, it’s highly recommended that you have Gear Lever preinstalled. This way, when you first launch Affinity-3.0.2-x86_64.AppImage, it will hijack it, allowing you to create a menu entry and to move the binary under ~/AppImages.

Of course, Affinity works, but I could never understand how people can learn the workflow of Adobe and similar graphics software! It would take me decades to get around such shit!

Moreover, I don’t even understand what kind of Affinity this thing offered by Canva is! In versions 1 and 2 (by Serif), there were Affinity Photo, Affinity Designer, and Affinity Publisher, to match Adobe’s obvious counterparts. But now, this unified crap might turn people away. OK, there are now several “Studios”: Pixel, Vector, Layout, and… AI. Fuck.

❻ Some of you might remember that I once complained that openSUSE Leap 16.0 doesn’t offer Live ISOs (click here on Alternative Downloads). But openSUSE Tumbleweed does (again, here on Alternative Downloads). And openSUSE-Tumbleweed-XFCE-Live-x86_64-Current.iso only has 1.0 GiB!

I needed to try it.

Not that bad for 1 GiB. But the first bug that I noticed was in YaST2. The last column in any list was partially out of the reserved area, thus triggering a horizontal scrollbar. Trying to resize it (or any column!) wouldn’t stick: the column would expand right away! The only fix is to double-click a column’s right border so that it fits to content. Thus, the total width decreases and the horizontal scrollbar disappears. But this is only temporary: maximize and unmaximize the window, and it will forget that it can fit! And every single window that has a multi-column list will start with a horizontal scrollbar because of this retarded behavior of the last column! Deutsche Qualität.

Remember Fedora XFCE trying to install yaru-theme? In openSUSE, what you need to install is xfwm4-theme-yaru:

Here, too, it installs too much, but less crap:

Only an icon is missing:

Like I did before, I wanted to install Iosevka Fixed or, with ligatures, Iosekva Term, for use as the default monospaced font. But openSUSE has a strange way of showing the Iosevka fonts (that was before applying a Yaru theme):

The Iosevka fonts are actually hundreds of ZIP files per release, but each release starts with a “Quick Picker” header that includes the basic fonts, not the SS01 to SS18 variations!

This isn’t optimal either, as Iosevka Term has ligatures, and I prefer Iosevka Fixed, but just replace “Term” with “Fixed” in the URL and you’ll get the other font!

In openSUSE, once I ignored the “ss01” to “ss18” variations, and also “aile/curly/etoile”… I was left with 2 choices:

  • iosevka-fonts, a 386 MB package
  • iosevka-slab-fonts, a 402 MB package

I installed the first one, which obviously brought all the “ss” crap!

But also the font I was looking for!

Damned openSUSE, with its stupid packages! OTOH, openSUSE’s repos include inter-fonts (12.6 MB), so I installed it:

While exploring what software is available in the default repos, without the need to enable extra repos (something that’s a mess, with inconsistent use of colons in a repo’s path), I noticed that fsearch was not available!

Options include: Flathub, Snapcraft (yes, snaps can be enabled!), and several 3rd-party repos hosted by SUSE, but I couldn’t access any of them (zzndb, thunderbird9) because software.opensuse.org was down!

Status:

Forums:

Reddit:

openSUSE is dying. That said, and not understanding a thing of their mess, I discovered that the same repos can be found on tracker.opensuse.org, whatever that was supposed to be. And both zzndb and thunderbird9 have fsearch-0.2.3-1.54.x86_64.rpm.

But openSUSE is anything but inspiring confidence.

❼ Back to Fedora: Using ACLs on Fedora Like a Pro (Because sudo is for Noobs).

I always hated ACLs because of… Microsoft Windows and Active Directory. But  traditional Unix/Linux permissions (rwx), distinct for Owner, Group, and Others, can’t cover modern use scenarios. Access Control Lists are actually a versatile concept.

Fedora and RHEL enable ACL support by default on most filesystems; in contrast, while ACL support is available in Debian and Ubuntu, it’s not enabled by default. Here’s a quick chat with Mistral about ACLs, SELinux, and AppArmor.

❽ Oh, but maybe FreeBSD is the solution!

Naïvely, I ran freebsd-update -r 15.0-STABLE install in my regular FreeBSD desktop. That was a mistake.

In principle freebsd-update is the way to update the installed system. In principle it works really smoothly, from binary release to binary release: it has a good manpage which tells you you can go from release to release. My FreeBSD laptop (which also runs Fedora 42 as more of a daily-driver OS) was running 14.3, so:

  • freebsd-update -r 15.0-RELEASE upgrade Fetch all the things.
  • freebsd-update -r 15.0-RELEASE install Install the new stuff.
  • reboot Smooth sailing.

That is the naive and optimistic thing to do.

If you spotted “make sure you read the announcement and release notes” in the manpage, good for you.

After the reboot, I was dropped into a shell where nearly every command I typed in resulted in

ld-elf.so.1: Shared object "libsys.so.7" not found, required by "libc.so.7"

❾ Soplos Linux is based on Debian testing, so I wanted to try their XFCE flavor. The strange thing is that Soplos.org correctly links to version 1.0.3 Beta (Xfce 4.22), whereas SoplosLinux.com links to version 1.0.1 Beta (Xfce 4.20.1). Either way, the real deal is on SourceForge.

In the live session:

  • Step 1: Install Synaptic.
  • Step 2: Synaptic isn’t anywhere in path. (It can’t be launched from the menu either.)

Good grief! And XFCE is still reported as 4.20 by their tool.

❿ Bluecurve GTK 3/4 (“Red Hat Bluecurve theme ported over to GTK 3/4. Designed for the MATE and Xfce desktop environments.”) is quite nice, with several color options, if nostalgia is something you experience. It also includes the Bluecurve icon and cursor set, and the Luxi font family. There is an install.sh (for the current user) inside bluecurve-3.0.0.tar.gz.

On RAM prices because of the AI madness

Friday evening, I was thinking about how the AI demand impacted and will continue to raise RAM prices. From what I remembered to have read:

  • Samsung, SK Hynix, and Micron control 95% of DRAM production and have shifted capacity to High Bandwidth Memory for AI data centers.
  • SK Hynix has sold out all 2026 production.
  • Micron can only meet two-thirds of customer demand.
  • Samsung raised some memory prices by 60%.
  • TrendForce expects average DRAM memory prices to rise 50-55% in Q1 2026 versus Q4 2025.
  • DDR5 RAM kits have already seen price increases between +46% and +138% between September and November 2025.
  • PC/laptop OEMs have been absorbing costs and stockpiling inventory since late 2025, but that buffer is running out now. IDC said that retail prices for smartphones and PCs will rise 10-20% or more in the first half of 2026.
  • Russian press announced that in March or April, the new smartphone shipments from China will bring a 10-30% raise in prices.

Meanwhile, new press articles caught my eye. The Reg: DRAM prices expected to double in Q1 as AI ambitions push memory fabs to their limit.

TrendForce this week revised its estimates with analysts now predicting DRAM contract pricing will surge by 90–95 percent QoQ, while NAND prices are expected to increase by 55–60 percent during the current quarter.

While AI demand is largely to blame, TrendForce notes that higher-than-expected PC shipments in the fourth quarter of 2025 further exacerbated shortages.

As we’ve previously reported, OEMs like Dell and HP tend to purchase memory in bulk about a year in advance of demand. If you noticed OEM pre-build pricing holding steady as standalone memory kits tripled in price, this is part of the reason why. But as inventories begin to draw down, and OEMs begin to restock, expect to see system prices climb.

TrendForce now expects PC DRAM to roughly double in price from the holiday quarter. And the firm forecasts similarly steep increases for LPDDR memory used in notebooks and other soldered-RAM systems, as well as in smartphones. TrendForce predicts pricing on LPDDR4x and LPDDR5x memory to increase by roughly 90 percent QoQ, the “steepest increases in their history.”

Then, in the NYT: How the A.I. Boom Could Push Up the Price of Your Next PC:

Since late summer, Mr. Reeves has been grappling with a tripling of the cost of memory chips, pushing Falcon Northwest to raise the price of some of its popular high-end computers to more than $7,000 from about $5,800.

“This isn’t a consumer-driven bubble,” Mr. Reeves, 55, said. “Nobody is expecting this to be a quick blip that’s going to be over with.”

The crunch in memory chips is the latest domino effect from the A.I. frenzy, which has upended Silicon Valley and lifted the fortunes of A.I. chip makers like Nvidia. The boom has now gone beyond A.I. chips to reach other components used to build gadgets, which could ultimately affect the prices of mass-market personal computers and smartphones, too.

Memory chip manufacturers can make more money selling expensive chip varieties to A.I. data centers than to the PC and smartphone companies that had long driven their revenue. As the chip makers focus on producing more for A.I. customers, their shipments of consumer-grade chips have slowed and their prices have surged — costs that could eventually be passed on to consumers.

TechInsights, a market research firm, predicts that higher costs for memory chips would raise the price of a typical PC by $119, or 23 percent, by this fall from the same period last year. Besides RAM, those figures include price increases for the chips called NAND flash memory, which provide long-term storage in computers and phones.

“The memory market is bananas,” Mike Howard, an analyst at TechInsights, said at a recent tech conference. “It’s very difficult to find any sort of fast relief.”

…

Micron, based in Boise, Idaho, is a microcosm for the changes. For 29 years, the company has sold different types of memory products to consumers and small businesses like Falcon Northwest.

But last month, Micron said it would discontinue that direct-to-consumer business, called Crucial, “to improve supply and support for our larger, strategic customers in faster-growing segments” such as A.I. chips.

…

“Nobody is getting everything they want, and we regret that,” said Manish Bhatia, Micron’s executive vice president of global operations. “We’ve had to make some difficult calls.”

Furious demand from A.I. is not just for holding data in more computers. Memory is also playing a more strategic role, with new kinds of chips determining how fast applications like chatbots deliver answers.

“The memory appetite for A.I. is so much larger,” said Michael Stewart, a chip industry veteran who is a managing partner at the Microsoft venture capital fund M12. “It’s utterly different than all the computing we did before.”

…

“Because our demand is so high, every factory, every HBM supplier is gearing up,” Jensen Huang, Nvidia’s chief executive, said at the CES trade show in Las Vegas, adding that the world would need more chip factories.

This all means tough choices for consumers, particularly the hobbyists who assemble or upgrade their own PCs. A typical PC kit containing plug-in modules of RAM chips that sold for around $105 in early September was $250 at the end of December, according to the website PCPartPicker.

Apple, Dell and other big companies often strike long-term agreements with memory suppliers that tend to moderate big price swings. Even so, the average retail price of a standard-configuration laptop PC jumped 7 percent in the two weeks ending Jan. 3, according to Circana, a market research firm.

…

And people shopping for a low-priced PC may get an unpleasant surprise.

“I definitely feel for the budget buyer,” he said.

Bananas!

Acer vs. Lenovo vs. RAM

Prices in Europe for low-end computers should seem unreasonably high to American readers, and they’re probably right: even if you remove the VAT (which is obviously already included in the end price), laptops are more expensive here than in the States. And so many cheap configurations aren’t available here at all.

But I was curious to see whether something interesting was still in stores before the apocalypse shipped from China hits our shelves.

👉 Here’s on Amazon Germany an Acer Aspire Go 15 (AG15-42P-R5ZD) Laptop, 15.6″ FHD IPS Display, AMD Ryzen 7 5825U, 32GB RAM, 1TB SSD, AMD Radeon Graphics, Windows 11, QWERTZ Keyboard, Silver, for €686.65 (including the German 19% VAT). For cheap shit, 32 GB of RAM is not bad (but it’s only DDR4), and 1 TB of PCIe Gen4 NVMe SSD is quite good!

But Acer is massively using MediaTek chips for Wi-Fi and BT, and Linux doesn’t always love some such chips. After the BT functionality in MT7663 stopped being supported in newer Linux kernels, I had to replace it with an Intel AX210. I hate opening laptops, but it can be done. This shouldn’t be the norm, though. So I needed to know what Wi-Fi/BT combo is used in this laptop. 

As AG15-42P-R5ZD is too specific a code (German keyboard and whatnot), what I searched for was the base code  AG15-42P. Note the “42P” part. Here’s an Italian disassembly of an Acer Aspire Go 15 AG15-42P. Right at minute 1:14, we see the truth: a MediaTek MT7921.

QUESTION: Is MT7921 better and more reliably supported than MT7663 by Linux, especially with regard to BT, not just Wi-Fi?

APPARENTLY, the answer is YES. The MT7921 seems to be significantly better supported than the MT7663, especially for Bluetooth. The mt7921e driver has been in the mainline kernel since around 5.12. The BT probably uses btusb and it certainly needs a firmware blob, but so far it hasn’t been dropped from newer kernels the way MT7663’s BT support was.

The caveat: MediaTek still isn’t Intel. So it’s not completely risk-free. (Sigh.)

👉 Lenovo doesn’t have anything dirt-cheap with more than 16 GB of RAM on Amazon Germany. This guy at €549.99 (including the German VAT) only has 16 GB: Lenovo IdeaPad Slim 3 15IRH10 (15.6 Inch WUXGA Display | Intel Core i5-13420H | 16GB RAM | 512GB SSD | Windows 11 Home | QWERTZ). But this generation of the IdeaPad Slim 3, namely “IRH10,” is worth exploring.

Since I can “hop” between two countries, Germany and Romania (except that in Romania most keyboards are US ANSI), I searched for offers in my “country of origin,” despite its higher VAT (21% vs. 19%).

On the beefy side (and cheap), only one offer: 

On the cheaper side, but with only 16 GB of RAM, which is a minimum these days (note that the prices are volatile):

The AMD-based ones are not “IRH10” but “ARP10”:

I’ve seen the last 4 laptops in Altex AFI Brașov, all having Linux Mint installed on them! Here’s the smallest one, the IdeaPad Slim 5 13ARP10:

The Wi-Fi and BT seem to be working (how about the webcam?)

The RAM is faster in cheap Lenovos than in cheap Acers, but even LPDDR5 isn’t always at 5600 MT/s.

But what Wi-Fi/BT combos are such laptops using? There is this Lenovo IdeaPad Slim 3 (15IRH10) Disassembly & Teardown video, paired with a detailed article.

There are two bad news here.

The Wi-Fi/BT combo is Intel AX203. While the Intel AX203 has good but recent Linux support through the standard iwlwifi driver, there are some caveats. Linux 6.14+ is generally required, though Intel’s documentation mentions 6.18+ for optimal compatibility. Funny thing, I wasn’t aware that even my AX210 has the same official recommendations, although I know from Xebian (“trixie” not “unstable”) that it works just fine with Linux 6.12. Maybe it only works in Wi-Fi 6, not Wi-Fi 6E mode.

But there’s a key difference between AX200/AX210 and AX201/AX211/AX203. Those ending in “1”or “3” require some hardware extensions and extra firmware to support the Intel CNVio2 protocol. I don’t even know what this shit is, but OK, the extensions are there as long as the laptop comes with an AX203. The problem is that I’m not sure I trust the Linux kernel and the distros regarding the AX203.

The second bad news is even worse. The teardown shows this SSD: WD SN5000S 512 GB PYRITE SDEPMSJ-512G. If you didn’t notice, the SDEPMSJ-512G = M.2 2242, which is difficult to find. The 2242 variant is essentially “unobtainium” for retail consumers. For consumers, this SSD exists as:

So you have the standard M.2 2280, the much shorter M.2 2230, but not the intermediate M.2 2242!

I asked Kimi to search for possible replacements. What can one do if such an SSD dies? Or, as long as these laptops come with  two Gen 4 2242 M.2 slots, where could I find a second SSD? And it was Kimi who told me this is “unobtainium”! The 2242 form factor is the “phantom” SKU that WD (SanDisk, actually) primarily sells to OEMs like Lenovo, not consumers. One can generally only find specialty, expensive 2242 retail SSDs, unless ordered from Lenovo. Or, maybe, older and slower PCIe 3.0 x4 instead of PCIe 4.0 x4. Still too expensive in Europe for what they are. What a PITA!

On the other hand, the laptop reviewed and disassembled is a strange animal. I knew that most IdeaPads come with soldered RAM (LPDDR5), but this one also has a DDR5 SO-DIMM slot in addition to the soldered 8 GB! So this model supports up to 40 GB of DDR5 at 4800 MT/s (8 + 32 GB)! From the video, it comes with 8 + 8 GB, so the slot isn’t free.

However, the cheap models listed above are limited to 24 GB, and those coming with 16 GB have 8 GB soldered + 8 GB in the slot. Not an optimal design by any means. But Lenovo isn’t tinkering-friendly in many regards. For instance, here I listed some ThinkPads that come with only one M.2 2280 NVMe slot. Now I learned that IdeaPads come with two M.2 2242 NVMe slots, which doesn’t help much because they’re shorter than the most common NVMe SSDs!

I am not thrilled. Cheap laptops with 24 GB or 32 GB DDR5 seem to have mostly gone. This is the end of cheap hardware. The AI apocalypse is coming!

Oh, and I forgot to mention that external SSDs have already either disappeared from stores or gotten more expensive!

On mental retardation regarding AI

We’re in 2026, and the press got frantic about something that happened in the spring of 2023!

Times of India: ‘Google model translated Bengali without being trained for it’: Sundar Pichai’s old clip on AI behaviour goes viral:

A short clip from a 2023 interview featuring Sundar Pichai has resurfaced online, reigniting debate about how modern artificial intelligence systems develop unexpected capabilities. In the clip, the Google chief executive describes a language model that was able to translate Bengali accurately after minimal prompting, despite not being explicitly trained for that task. Pichai also acknowledged that parts of such systems function as a “black box”, meaning their internal decision-making is not always fully understood, even by the engineers who build them. The resurfacing has sparked fresh discussion around AI transparency, control, and public trust.

What Sundar Pichai said in the original interview

The comments were made during a 2023 appearance on 60 Minutes, where Pichai and other Google executives discussed the rapid progress of large language models. He cited the Bengali example to illustrate how AI systems can display abilities that engineers did not directly programme or anticipate.The model was not learning Bengali from scratch in real time. Instead, it had already been trained on vast multilingual datasets during its initial development.

What surprised researchers was how quickly the translation ability appeared once the right prompts were given. This phenomenon, where new skills seem to emerge suddenly as models scale, is commonly referred to as “emergent behaviour” in AI research.

This an ‘emergent ability’

In large language models, certain capabilities do not improve gradually. Instead, they can appear abruptly once a model reaches a particular size or complexity. Translation, reasoning, and few-shot learning are among the skills known to surface this way. While researchers can measure when these abilities appear, explaining precisely why they emerge at that moment remains an active area of study.

Compare to an article from 2023: Google Surprised When Experimental AI Learns Language It Was Never Trained On; and a video from back then:

Humankind is utterly retarded.

Regarding the 2023 interview featuring Sundar Pichai, in which he talked about an LLM that was able to “learn,” or at least to translate Bengali, despite not being explicitly trained for that task, I have a non-technical theory off the top of my head.

Simplified, I take it that the said LLM, having been fed a humongous amount of documents, including some in Bengali, was able to “make correlations” or to “know without knowing” what something written in Bengali corresponds to in English. It surely helped to have “digested” non-fiction materials in Bengali, not just fiction. Scientific papers with drawings and pictures, for instance.

I would bring two comparisons with humans to support my theory that such an “emergent ability” is entirely normal.

Firstly, a baby doesn’t know any language, and there’s no way to explain to them grammar or definitions. Yet, after being exposed to lots of uttered words, babies make correlations and “deduce” the meaning of basic words. (Mind you, this is not an easy task, because the same word is pronounced differently by different persons, so the baby has to classify similar utterances as representing the same concept. As a side note, when adults from certain cultures have difficulties in differentiating the sounds for “b” and “p,” this is because in their mother tongue there is a unique sound situated between “b” and “p,” closer to one of them, but to such people “b” and “p” classify as the same sound.)

Secondly, in the 1980s, I used to read the Soviet magazine Радио. I only knew a few Russian words, but I had this tactic: by examining the schematic diagrams, I gathered enough information about the functioning of a proposed device, that this helped in understanding what the text wanted to convey about the building and adjusting of the device. It was not “learning Russian,” yet it was an “emergent ability” helped by the drawings and schematics. But I could only understand some technical texts in Russian, and common wording used in such contexts.

Mistral agrees with me. There is nothing to worry about. Oh, and Sundar Pichai is a highly overpaid retard.