FreeBSD is well situated

DanielBowman · Monday at 4:07 PM

Agentic "AI" seems to be the (near) future standard. No more static DE GUI, just cli, voice and a few dynamically generated knobs.

What better system to run the agents on than POSIX compliant and well documented stable system?

Wayland and it's somewhat tightened security might actually be a problem for Agents...yey XLibre.

Permissive license to load it on a box on wheels with some camera and sell it as your personal assistant.

I am actually hopeful for FreeBSD...if done right. It is positioned to be the best option for semi-autonomous computers.

MG · Monday at 4:20 PM

I don't understand the goal. Something must happen while obscured to the operator? Why do "agents" need to control everything somehow more than a logged in root? All can be done by only a program, right?

Zare · Monday at 4:45 PM

AlfredoLlaquet · Monday at 4:47 PM

I generally don't like memes, but this one is spot on. I'm roaring with laughter.

cracauer@ · Monday at 5:00 PM

Would be more fun with official CUDA support.

drhowarddrfine · Monday at 7:59 PM

This is also the year of FreeBSD on the desktop!

DanielBowman · 2026-02-10T18:17:19+0000

drhowarddrfine said:
This is also the year of FreeBSD on the desktop!

IMO Desktop as we know it is done for the most part. It is going to be just a canvas.

DanielBowman · 2026-02-10T18:19:44+0000

cracauer@ said:
Would be more fun with official CUDA support.

Yeah, that might be a problem. Hopefully someone (AMD, RISC-V, Intel if resurrected ) will end NVIDIA close source era.

MG · 2026-02-10T18:26:32+0000

DanielBowman said:
IMO Desktop as we know it is done for the most part. It is going to be just a canvas.

Everything works except 3d accellerated graphics. Gamers should cancel it. We just need fast enough computers to do the same.

AlfredoLlaquet · 2026-02-10T18:38:33+0000

The OP is too optimistic. We're still far away from the widespread adoption of "AI agents" because they are based on LLMs, and LLMs, even the one that use a reasoning loop, are very very very unreliable. To be precise, they invent things; they made things up... A LOT. You cannot even get them to be reliable customer service operators. In practice, no one is using AI in an industrial scale. For the moment, it's a nice shiny toy and I love to play with it.

MG · 2026-02-10T18:56:46+0000

AlfredoLlaquet said:
The OP is too optimistic. We're still far away from the widespread adoption of "AI agents" because they are based on LLMs, and LLMs, even the one that use a reasoning loop, are very very very unreliable. To be precise, they invent things; they made things up... A LOT. You cannot even get them to be reliable customer service operators. In practice, no one is using AI in an industrial scale. For the moment, it's a nice shiny toy and I love to play with it.

I think it has a commercial problem. If your prompt-service actually provides constructive information that has a market value, you're giving away profit.

cracauer@ · 2026-02-10T19:11:02+0000

MG said:
Everything works except 3d accellerated graphics. Gamers should cancel it. We just need fast enough computers to do the same.

3D graphics work on FreeBSD.

It is compute on GPUs for NVidia hardware (CUDA) that is a problem.

MG · 2026-02-10T19:44:53+0000

cracauer@ said:
3D graphics work on FreeBSD.

It is compute on GPUs for NVidia hardware (CUDA) that is a problem.

How can I install and play Warcraft 3 Trozen Throne in a vm or emulator on FreeBSD? It needs an old Nvidia driver and DirectX 9. Extra nice would be if a method exists that can be reproduced without needing something proprietary. It's like 25 years old...

DanielBowman · 2026-02-10T22:28:53+0000

AlfredoLlaquet said:
The OP is too optimistic. We're still far away from the widespread adoption of "AI agents" because they are based on LLMs, and LLMs, even the one that use a reasoning loop, are very very very unreliable. To be precise, they invent things; they made things up... A LOT. You cannot even get them to be reliable customer service operators. In practice, no one is using AI in an industrial scale. For the moment, it's a nice shiny toy and I love to play with it.

It is still way "smarter" and certainly faster (and cheaper) than average human. 80% of humans need just 3 applications on a computer - email, browser, Word (rest 20% also Excel).
Which is all basically just text | pic in & out. All with very few basic scenarios. LLM is smart enough to cover those. For more complex tasks, even few times bungled help from LLM is still more efficient than studying the menus, help or manuals only to start again when the SW changes.
And the LLMs or rather agents are still improving. Where they lose in logic they gain in experience. You can see that with coding ability. They lied, hallucinated often and now they create very complex programs, do test loops etc. Even without hard reasoning or logic - they do what most of us did...copy and paste and glue together a code from somewhere. And agents can do that from 100 different angles, give feedback and correct itself.

Espionage724 · 2026-02-10T22:35:55+0000

MG said:
How can I install and play Warcraft 3 Trozen Throne in a vm or emulator on FreeBSD?

I have 1.27 notes! It ran easy with Wine on 14.2 (15+ pkg32-old notes)

-opengl probably could be used with MESA_OVERRIDE Zink for Vulkan (if OGL by itself might be non-ideal)
D3D mode was unstable on Intel UHD 630 (not sure a benefit over OGL, but could probably work with DXVK)
I use 1.27a installers (not Reforged) without requiring Bnet client (updating past that to like 1.30 hooks Bnet; versions info)
There's ideal versions of WC3 (1, 2)

cracauer@ · 2026-02-10T22:50:36+0000

Espionage724 said:
I have 1.27 notes! It ran easy with Wine on 14.2 (15+ pkg32-old notes)

-opengl probably could be used with MESA_OVERRIDE Zink for Vulkan (if OGL by itself might be non-ideal)

D3D mode was unstable on Intel UHD 630 (not sure a benefit over OGL, but could probably work with DXVK)

I use 1.27a installers (not Reforged) without requiring Bnet client (updating past that to like 1.30 hooks Bnet; versions info)

There's ideal versions of WC3 (1, 2)

I think IMG's problem is that he has a too old NVidia card. And since we have no control of the binary driver...

MG · 2026-02-10T22:57:12+0000

Espionage724 said:
I have 1.27 notes! It ran easy with Wine on 14.2 (15+ pkg32-old notes)

I had a few old games working with Wine but its on the wrong level. Why can't Virtualbox or Qemu have the required virtual logic that enables Wine to do it? (different for every game because it depends on dll versions)

ralphbsz · 2026-02-10T23:04:50+0000

DanielBowman said:
80% of humans need just 3 applications on a computer - email, browser, Word (rest 20% also Excel).

Actually, I would contend that 99% of humans only need a browser. It's perfectly feasible to do documents (a.k.a. Word), spreadsheets (a.k.a. Excel), and mail within a browser. Same with e-mail. The remaining 1% is gaming, graphics (not CPU!) intensive image processing, and some CAD that is too complex to convert to a browser.

Another application for desktop/laptop computers is as user interfaces for complex machinery. For example, if you are using a high-end 3D scanner or a high-end CMM (coordinate measuring machine), you typically need custom software, and attach a computer to the machine. That custom software tends to only exist for Windows. Similar problem with software to initialize/manage/configure industrial machinery (such as PLCs), again that tends to only run on Windows. Another example is LabView, but that at least exists for Mac and Linux.

DanielBowman said:
And the LLMs or rather agents are still improving. ... And agents can do that from 100 different angles, give feedback and correct itself.

I love your optimism about both AI and agents computing. I don't share it at all.

But even if both fields succeed, most of the AI inference and use of agents will be on servers, which will mostly be in cloud data centers or corporate computer rooms, not on people's laptops. The energy efficiency, RAM bandwidth, compute power, and parallelism just aren't there on a device you can carry around with one or two hands.

drhowarddrfine · 2026-02-10T23:58:41+0000

Since LLMs get all their info from the internet, I wonder if the first major attack will be on them by evil forces feeding them false information. I wonder if the attack will be constructed using AI itself.

DanielBowman · 2026-02-10T23:59:11+0000

ralphbsz said:
Actually, I would contend that 99% of humans only need a browser. It's perfectly feasible to do documents (a.k.a. Word), spreadsheets (a.k.a. Excel), and mail within a browser. Same with e-mail. The remaining 1% is gaming, graphics (not CPU!) intensive image processing, and some CAD that is too complex to convert to a browser.

Another application for desktop/laptop computers is as user interfaces for complex machinery. For example, if you are using a high-end 3D scanner or a high-end CMM (coordinate measuring machine), you typically need custom software, and attach a computer to the machine. That custom software tends to only exist for Windows. Similar problem with software to initialize/manage/configure industrial machinery (such as PLCs), again that tends to only run on Windows. Another example is LabView, but that at least exists for Mac and Linux.

I love your optimism about both AI and agents computing. I don't share it at all.

But even if both fields succeed, most of the AI inference and use of agents will be on servers, which will mostly be in cloud data centers or corporate computer rooms, not on people's laptops. The energy efficiency, RAM bandwidth, compute power, and parallelism just aren't there on a device you can carry around with one or two hands.

Browser is enough, yes. But I spoke about it regarding GUI - menus etc.. Standard browsing is just clicking and back button.
With AI you don't need Menus. GUI can be dynamic, generated according to current workflow.
This does not need Huge LLMs somewhere on a server even now - on decent HW that is. 14b (not specific for this task) model can probably do it and much faster without the Internet/GPU Farm latency. Especially if the model is trained for it.
Even now, you can buy google's Coral USB and make Rapsberry pie do Vision to LLM.

Why ask ChatGPT how to chage heading styles in Word if you can have miniGPT do it for you. Same with syntax in command line or editing config files.

Those complex CAD etc. SW will also use it. Maybe it will even force them to open their API, so any current model can be utilized on them. Once they do not have to compile for specific OS/DE they will be happier for it.
Microsoft knows this. That is why they named O365 a Copilot - because eventually it will be all that remains - you tell it what you want it to look like and it will assemble it.

From all the current OSes, FreeBSD shines - Windows is lacking in command line. Mac is closed source and their main advantage is clean(ish) GUI and this will be irrelevant with dynamic LLM UI. Linux is open but a mess.

Espionage724 · 2026-02-11T00:08:04+0000

DanielBowman said:
Windows is lacking in command line.

Meh, I create wild lines fine

(I'd be interested in seeing more unique msbuild/clang use anywhere else)

Code:

"%ProgramFiles%\CMake\bin\cmake.exe" -S "%UserProfile%\Projects\TrinityCore-335\src" -B "%UserProfile%\Projects\TrinityCore-335\build" -G "Visual Studio 17 2022" --fresh -Wno-dev -DBOOST_ROOT="%SystemDrive%/local/boost_1_90_0" -DMYSQL_INCLUDE_DIR="%ProgramFiles%\MariaDB 12.2\include\mysql" -DMYSQL_LIBRARY="%ProgramFiles%\MariaDB 12.2\lib\libmariadb.lib" -DWITHOUT_METRICS="1" -DTOOLS="1" -T "ClangCL" -DBoost_COMPILER="-vc143" -DCMAKE_CXX_FLAGS="/EHs /arch:AVX2 /favor:INTEL64 /Gz /Oy /GS- /guard:cf- /W0 -Wno-everything"

It's a little more boring FreeBSD (a lot of that stuff is auto-detected and probably safe to trust with the OS pkgs/deps over-time):

Code:

rm -Rf ~/'Projects/TrinityCore-335/build' && mkdir -p ~/'Projects/TrinityCore-335/build' ~/'Projects/TrinityCore-335/run' && cd ~/'Projects/TrinityCore-335/build' && cmake ~/'Projects/TrinityCore-335/src' -DCMAKE_INSTALL_PREFIX=~/'Projects/TrinityCore-335/run' -DCMAKE_BUILD_TYPE='Release' -DWITHOUT_METRICS='1' -DTOOLS='1' && make -j "$(nproc)" install