Forgot your password?
typodupeerror
Windows AI Iphone Programming

AI-Assisted Coders Run Windows on iPhone Duo, Build a Floppy Disk-Based AI Chatbot for MS-DOS (xda-developers.com) 44

The blog XDA Developers reports that a Claude enthusiast used the AI assistant to make a language model that runs off a floppy disk. "And you can stick it on a DOS computer to boot."

The language models "runs at a respectable speed, given that it's running off a CPU that was designed decades before LLMs were even a thing." Over on the ClaudeAI subreddit, user wollymonjak showed off this really impressive feat... [I]t's not going to be a particularly powerful AI; in fact, wollymonjak notes that if you ask something too complex or too long, it starts spitting out nonsense. But who cares about that when you're using a language model on DOS?
The same blog also notes that "the iPhone Duo hasn't been out for too long, and someone has already gotten Windows 11 on it." As spotted by TechSpot, this cool feat was performed by Ammaar Reshi over on X. Their project is called WindowsiOS, and its goal is to get Windows 11 running on an iPhone Duo without jailbreaking. Usually, when we cover a project coded up by AI, Claude is the one that called the shots; this time, Reshi opted to use Gemini 4 Carbon, Google's flagship model.

AI-Assisted Coders Run Windows on iPhone Duo, Build a Floppy Disk-Based AI Chatbot for MS-DOS

Comments Filter:
  • Yes (Score:5, Funny)

    by BitterEpic ( 10503015 ) on Monday November 16, 2026 @07:51AM (#66426620) Homepage

    These are exactly the kind of articles I want to see (sarcasm.) Maybe they can gemini code something to remove the unremovable AI buttons all over the place. Or maybe gemini code a robot that will kill Altman or Pichai./p>

    Those would make me feel happy after my lost 3 hours of coding with Opus constantly lying to me at work.

    • Re:Yes (Score:4, Informative)

      by nospam007 ( 722110 ) * on Monday November 16, 2026 @09:23AM (#66426766)

      Yes it could.
      Dedicated Extensions:
      Add-ons like AI Content Shield (available for both Firefox and Chrome) strip out Google AI Overviews, AI Mode tabs, Bing/DuckDuckGo AI summaries, YouTube AI video summaries, and AI prompts across major platforms.

      Similarly, Search Cleaner strips AI prompts from search engines.

        uBlock Origin (Custom Filters): Rather than installing yet another extension, most power users simply add custom cosmetic filter lists (often shared on GitHub and Reddit) to uBlock Origin.
      Since uBlock Origin already parses the DOM on every page load, adding CSS selectors to hide .ai-summary, [aria-label*="Ask AI"], or Google's AI Overview containers does the exact same job with zero extra overhead.

      Browser-Level Kill Switches:

      For the browser's own built-in AI buttons, Firefox (version 148+) added a single "Block AI enhancements" master toggle under Settings > AI Controls that removes sidebar chatbots, AI tab-group suggestions, and link-preview summaries in one click. (In Chrome, you still have to hunt down flags or registry policies.)

      Any modern LLM can write a working WebExtension (manifest.json + a 40-line content.js using a MutationObserver) in seconds. It simply scans the page for , , or role="button" elements whose aria-label, title, or text matches keywords like "Ask AI", "Summarize with AI", or "Copilot", and sets display: none.

      The only reason a generic AI-written script isn't 100% bulletproof across every site simultaneously is collateral damage:

      False positives:
      A naive keyword-matching script will accidentally hide the "Ask AI" tag on a Slashdot headline, a button on an AI news site, or a legitimate "Email" / "Detail" button if a developer wasn't careful with regex word boundaries (\bAI\b).

      Obfuscated class names & Shadow DOM: Sites like Google, LinkedIn, and X constantly randomize CSS classes (class="x1i10hfl...") and hide widgets inside closed Shadow DOMs, meaning a universal heuristic script has to keep being updated—which is essentially turning it into an ad-blocker filter list anyway.

    • What's worse than trusting an AI oracle too much?

      Trusting an AI oracle that knows nothing but has a high priority to flatter you and tell you whatever it "thinks" you want to hear.

      Worst: The AI oracle has hidden "motives" coming from (1) a sociopath with infinite greed or (2) a fascist government or (3) both.

      Scare-quoted words may not mean what you think they mean. But you can ask your favorite AI to interpret them for you.

    • I use a TamperMonkey script to zap the Alexa side window that pops up on Amazon's website. It doesn't quite block enough though because the website is still really slow on one of my lower-end PCs. Makes me not to want to use Amazon. Walmart does free delivery on orders above a small amount, no membership fee, and has a less invasive site. There are lots of options although none that cover the same breadth as Amazon.

  • Insanely small (Score:5, Interesting)

    by Rei ( 128717 ) on Monday November 16, 2026 @07:53AM (#66426622) Homepage

    A "tiny" LLM is generally considered to be about 1B params or less. This one is 366k params.

    Of course, it seems to be immensely narrowly trained: "Trained on about 3,500 facts, each asked in many wordings Outside those it makes things up confidently"

    What would be more interesting, IMHO, would be to try to train a RAG model. Something that is trained on *no* facts, merely on converting your statements into search terms, and turning search results back into answers. I wonder how small you could get that? I've often wondered how far back we technically could have run an actually-useful LLM if we were to have had the weights and inference engine.

    • Re: (Score:3, Funny)

      by OrangAsm ( 678078 )
      Why not 640k? That should be enough for anyone.
    • Re:Insanely small (Score:5, Interesting)

      by Rei ( 128717 ) on Monday November 16, 2026 @08:17AM (#66426662) Homepage

      ED: Looking into it, the state of the art for actually-useful RAG is still about 60-100M params, so still way off from what you can fit on a floppy disk. Quantized, you're looking at maybe 25 floppy disks, 30-50MB at 4-bit compressed.

      For running it comfortably, a '96-98 era Pentium MMX / Pentium II fully-in-ram with 64-128MB ram and a 1-4GB hard drive should be easy and run at a workable (go brew a cup of coffee ;) ) 1 tok/s out.

      For at least being able to fit it in RAM, you're looking at the 1987-1990 era - a Mac II with the FDHD could support up to 64 MB of RAM, and a 386DX could handle 32-64MB on some motherboards. It would have cost tens of thousands of dollars and taken like half an hour per answer.

      If you're really damned patient and willing to stream from disk, we can go back to 1984-1986 (IBM PC AT / XT + 40MB HDD). It could just keep the current hidden state and a tiny 640k weight buffer in RAM and stream the model layer by layer. Prefilling a tiny 300 token passage would take like 2 days, and a 40 token answer would take like 5 hours. On a 4.77 MHz 8088 PC, you're talking maybe 1-2 weeks per query.

      And if you're a masochist... technically you could stream straight off of manually-swapped floppy disks! Then you could run it on a 1977-1982 Apple IIe or Commodore 64 ;) It'd take ~200x 5 1/4" disks. Of course, merely the stage of generating a 30 word answer would take around a month and six thousand floppy swaps. ;) Once every ~7 minutes.

  • They will be phasing out floppy disks from air traffic control systems.

    Of course, before the last one goes, one enterprising techie will test this chatbot:

    Would you like to play a game?

    • by Rei ( 128717 )

      The funny thing is, if we could send the weights and inference back in time, a 1983 Cray X-MP/2 running a 300M–400M total / ~35M–40M active param MoE - probably similar to the reasoning capability of the computer in Wargames - could probably output 1-3 tok/s (or make an equivalent number of reasoned decisions per second with a Jev-like decision head). In Wargames, the computer seems to output at about 3 tok/s [youtu.be].

      Who knew that that movie was technologically plausible? ;)

      • by Rei ( 128717 ) on Monday November 16, 2026 @08:53AM (#66426728) Homepage

        Basically, the Cray X-MP/2 had 16-32 MB of fast RAM and a then-absurd ~5 GB/s memory bandwidth and a solid state "disk" (actually MOS RAM cache) with 256MB-1GB of storage and a 100-1000 MB/s bandwidth / sub-25-microsecond latency. The DD-39 / DD-49 disk arrays could hold 1,2GB per drive, up to 38GB, 6-10MB/s unstriped or ~50MB/s striped. Honestly, you could readily go with a far larger number of total params, so long as you kept the active params small enough to fit in RAM. An architecture like Freetoken could significantly improve performance as well. Performance would be limited by the 400 MFLOPS of compute, not the bandwidth (really, apart from the fact that all math was 64 bit float, it wasn't a bad architecture for inference!). You'd want to do token batching and maybe add a speculative decoding model to boost performance further.

        Of course, you could never train a model like that on the Cray!

    • I remember (faintly, It's been a while) the ATC simulator that actually ran on MS-DOS PCs. We were impressed with the remote controller's voice being simulated through the PC's bleep speaker. Speed was adequate as well.

  • Finally, a clear evidence that these tool, so called "Artificial Intelligence" are not intelligent...
  • re. this sentence "And you can stick it on a DOS computer to boot."

    Is this a boot floppy, meaning the writer made a clever pun, or is it not, meaning the writer is an ignorant moron ?
  • by hcs_$reboot ( 1536101 ) on Monday November 16, 2026 @08:41AM (#66426704)
    is like eating a McDonald's hamburger in a fine-dining restaurant.
  • by Pizza ( 87623 ) on Monday November 16, 2026 @09:25AM (#66426772) Homepage Journal

    ...Only a lot slower/

    (the MS-DOS port of Eliza, which is itself nearly 60 years old, could fit in 16K and run on the original 4.77MHz IBM PC)

    • by Rei ( 128717 )

      "Tell Me You've Never Used ELIZA Without Telling Me You've Never Used ELIZA"

  • by rsilvergun ( 571051 ) on Monday November 16, 2026 @09:48AM (#66426828)
    These are fun projects to build but the point is building them. It's like Legos. Who the hell would want to buy a pre-built Lego?

    If it's something you're going to use I get it and I understand buying an action figure that's ready to go because you're going to use it as a decoration. But slop coding useless hobby projects just seems pointless...
    • A lot of people do like Playmobil (the closest thing to a pre-built Lego I guess).

      I guess those people also enjoy prompting and getting a fun result without too much effort.

      • Yeah but if you're not a kid anymore you are putting them on a shelf to look at.

        You don't really put software on a shelf to look at at least unless it's in a box.
    • by DarkOx ( 621550 )

      Honestly I think that is best part of this; you can play with the bits that actually interest you and let AI filling the gaps and/or do the parts that are not your bag.

      If you are into retro gaming or something and you want to run win95 on contemporary VM platform with a display driver that is better than basic SVGA/Vesa; have AI whip you up a driver. If you are into legacy Windows internals sure it is playing with assembled legos but most people are not, understating legacy Windows internals isn't very appl

    • by Rei ( 128717 )

      If you think you can just type "Make a LLM that can run on a floppy disk", leave your session sitting there for a while with zero interaction, and come back and it's done and distributed, I have some terrible news for you...

    • "Who the hell would want to buy a pre-built Lego?"

      -1, offtopic

      No one is saying it's now a solved problem ready to be productized, *except you*. They're demonstrating that a weird thing is possible.

  • The question used to be "Can it run Doom?" The new challenge is "If it ran Doom, can it run an LLM?"
    Pretty soon all sorts of things will start talking back to us, trained with wise-ass comments and flippant answers.

    • by Rei ( 128717 )

      I'm disappointed by how little my toaster talks, and I need to remedy that situation.

  • Well we already had ELIZA running on low-end hardware, well enough to fool some people.

    • "low-end hardware" ?!

      What ?! Why, it took all of my rather expensive Commodore Pet's 8,192 bytes of RAM to run Eliza from Ahl's More Basic Computer Games.
    • by Rei ( 128717 )

      I can only find a single example [mit.edu] in all of history of a person who was fooled by ELIZA, and only in a "he thought it was a remote-working employee taking the piss" way, not in a "he thought a computer was alive" way.

      This notion that ELIZA mass-fooled people is simply not in line with the actual reality we live in. Weizenbaum pushed it heavily in his later life because he was an anti-AI zealot (driven in part due to anger about the use of computers for targeting during the Vietnam War, but also because some

      • Best statement in the linked article: Weizenbaum would write of ELIZA, “What I had not realized is that extremely short exposures to a relatively simple computer program could induce powerful, delusional thinking in quite normal people.”
      • I can only find a single example [mit.edu] in all of history of a person who was fooled by ELIZA, and only in a "he thought it was a remote-working employee taking the piss" way, not in a "he thought a computer was alive" way.

        So in other words, ELIZA passed the Turing test, which is kind of a big deal for AI.

        • by Rei ( 128717 )

          The Turing Test has long been considered highly defective. It also doesn't mean "a person thinks the computer is alive". It means "the person thinks what they're talking to isn't a computer".

          If you wrote a program to write out random clusters of adjacent letters, and had people try to guess whether they were talking to a computer or a cat walking on a keyboard, a lot of them would guess wrongly that they were seeing a cat walking on a keyboard. That doesn't mean that any of them think computers are cats, t

  • ...in fact, wollymonjak notes that if you ask something too complex or too long, it starts spitting out nonsense. But who cares about that when you're using a language model on DOS?

    I don't need AI to vibe code software that doesn't run properly. My prior art goes back to 1982.

  • Dr. SBAITSO has entered the chat.

    https://www.youtube.com/watch?v=1lzJMJxTA7k
  • Perhaps some truly open, opensource, and free-as-in-beer AI software will damage the AI market, as in please make it.
    Right now, AI is owning everyone, everyone's privacy, everyone's creative content, everyone's RAM, everyone's electricity, everyone's future.
    • Have you not seen some of these local Qwen models And of course, as RAM speeds increase and CPUs get faster, AI will be running locally much faster than it does now..
  • Nothing new here.You can tell AI what hardware to target. I've had AI write stuff for a 486, no problem. People have Vibe coded demakes all the way back to the Sega Genesis. Which also throws the excuse that we can't make modern games on older hardware out the window. In an era where nobody can buy, you know, new hardware.

"There are some good people in it, but the orchestra as a whole is equivalent to a gang bent on destruction." -- John Cage, composer

Working...