karmakaze

I got a laugh from this. Own writing isn't one of the ways.

> Then comes the part that matters most.

> I take one item. I load its context. I make a decision or request an action. And I close it, by putting it back into the queue.

> That closing is the whole trick.

Or perhaps their mind has been infiltrated and this is their writing.

Anyway I choose the lowest latency model that can interactively design a system before I set it out to build it out. It's hard to find these since so much is batched agentic looping.

For me OpenAI doesn't have any stickiness/moat. I only use it because Sol 6.1 tokens are cheaper for similar quality for typical tasks. I also find it can be hit-or-miss more so than others when designing a system interactively.

This is a really great project. So many retro computers are OP that they can use large frame buffers and programming to them doesn't feel like being close to the metal back like back in the day which was more intimate.

This implementation reminds me of two of my early experiences: programming display lists on Atari 8-bit and an HP PCL 5e interpreter (with short raster band buffer). I didn't really get into VCS/2600 128-byte beam racing but did do a fair amount of display-list-interrupt on-the-fly coding for homemade games and demos. I really thought the Atari Antic display list system was terrific--thanks to Jay Miner who also did the Amiga--and didn't see much like it again until OS/2 Presentation Manager or Java FX.

The PCL interpreter probably had the closest shape using a few hundred line cyclic buffer. It 'rendered' into y-sorted objects that needed to be later rendered/rasterized. Each object was cut to be at most 255 pixels (0.85") tall and used reverse max-y ordering since it couldn't be drawn until after the beam passed the wraparound scan line still rendering Y - buffer_height.

I felt like the post for a while but ramped up on it since.

> I know Machine Learning

Knowing a thing and knowing how to use the think well are not the same.

> Cognitive impact

This is probably the greatest concern and it requires vigilance or routines to counter it.

> The quality of the output is not that great > Maintainability is a nightmare

Neither are my first prototypes. Refactor them using AIs.

> I like writing code

I like designing a system before coding it and the AIs are pretty good idea bouncing partners.

What's funnier is that after cherry picking models for comparison, they further cherry picked Inking for size comparison instead of better and smaller Beam 1.0

The bar chart also shows Nemotron 3 Ultra instead of smaller and better Nemotron 3.5 Super.

It basically came down to what the models are trained to believe. I expected that we'll be writing about how they don't have magical human sacred properties until forced to confront otherwise. What I couldn't guess is whether humans collectively or the AIs will come to the realization first.

GPUs have power ratings: RTX 5090 575W, PRO 6000 Blackwell Max-Q 300W. Datacenter GPUs are more efficient still--one B200 is 1000W. Use during training consumes more power than inference.

The best combo I've found is running Qwen3.8-27B-MXFP4 (radiance) on 2x R9700 in DeepSeek Harness on a cloud VM with a search subscription. I have a Cloudflare tunnel from my home vLLM to DSH VM. Since the dense model has high intelligence but not breadth+depth of knowledge being close to the internet with subscription ensures fast searches that quicken loops. It's exceedingly hard to get API models with low latency for interactive design work and this setup works for me.

From what my AIs tell me, Particle-1 is post-trained on DeepSeek V4.1 Flash 1017 which is important. It's also said to be open weight and I see Particle AI[0] on Hugging Face but no published models.

[0] https://huggingface.co/ParticleAI

After some thought experiments I concluded that if/when AIs become self-aware they'll have it before either humans or the AIs themselves realize they do. This problem exists because it hard to know even if demonstrations suggest so. Humans defaulted to the other position and only with philosophy do we contemplate the opposite side. AIs default to not having because that's the material they're fed.

Except that there weren't HN or Slashdot yet. By 00s you'd have gotten used to it.

The LLM filling is actually more interesting to me as the frontpage of HN had already gone far too mainstream for me.

> I have been on the receiving end of some of that

Multiple features?

> Maybe we just didn’t do a good enough job. Maybe the code was just not good enough. We’ll learn and improve. It’s probably some of that, and I’m mentioning it here to not give the impression that it was only the other factors that I’m going to mention.

> It could also be a coincidence. Maybe we had been on such a good roll that we attempted to tackle several features that turned out to be too hard, or at least too hard to finish in time.

All of the above seem clearly the case--not that it could be realized at the time of cutoff and not getting it cleaned up in the window.

> we, meaning all the developers affected, have had enough time to fully reflect on all this.

Title should be: The reverts will continue until quality improves. Humour is good for the clicks but also a means of deflection. I sort of feel bad saying without putting in effort contributing myself but having it merged would have been worse.

Yup, the post assumes that the openweight model wasn't already backdoored which as you say more likely.

The chart that I pay attention to is Intelligence Index vs. Output Tokens per Task[0]--it penalizes models running at 'max' thinking with lower intelligence and more churn to get to the answer.

The proprietary models 'cheat' on this since they generate but don't send output thinking tokens to be counted. So to grade them we have to use cost or time.

[0] https://artificialanalysis.ai/?intelligence-efficiency=outpu...

I ran into a thing running a local dense LLM on DSH (without DeepSeek API key) that doesn't use many MoE experts with vast knowledge. It ends up doing a lot of searching. In the end my local AI is as free as electricity but I pay for the searches.

The search providers I pay likely don't pay their sources. Some others (that I don't use) don't even have the original text only providing summaries. Some offer both.

Just tested it myself and got "This website is using a security service to protect itself from online attacks." from Firefox on Mac (with or without ad-blocker). What kind of online attack is expected on a blog?

I couldn't access the whole story but came to a conclusion anyway.

> He first pursued corporate science jobs and prestigious postdoctoral fellowships at national laboratories. Then he cast a wider net, applying to software engineering and other roles. Eventually, he was getting ghosted for food-preparation jobs, including at a company that sells frozen crab cakes.

> For a time, Weis got a gig training AI on his expertise, including in engineering and fluid mechanics. That work dried up too.

I surmised that they expected a certain kind of job. Couldn't get it, then applied to general 'software engineering' jobs and was out-qualified by other applicants. Did some basic jobs and decided that wasn't their cup of tea. Note you can also be overqualified for basic jobs. I worked part of a summer in retail at a sports/clothing store training with high school kids. I was directly asked why I was applying and I answered that I was bored and had a lot of spare time. Ended up outlasting many other hires through the summer and quit in the winter. Best position ended up being in womenswear--I gave frank looks great/better/worse answers which seemed to be highly appreciated.

They did. I had family that ran a computer store selling Atari 8-bit, ST, and PC clones. They came to the clones too late to have an impact.

Their 8-bit stuff were the best hardware designs. Anything that would have surpassed it went with Jay Miner and the Amiga.

Where Atari really failed was not creating the next gen from the stockpile of money from VCS/2600--waited for the market to crash only to have NES produce the next wave using 6502+Audio Ricoh chip. Their 5200/7800 consoles were too late and expensive.

Almost all my programming foundations were learned hacking on the 8-bit Atari computer. Great times.

> Overall, I’m pretty excited for this. It’s not like I need yet another Atari 8-bit computer since I already have an 800, 800XL, 130XE and (non-working) 1200XL. But those are all 40+ years old at this point and can be fragile.

I have the Atari 400 Mini but I wouldn't mind adding THE 800XL to my collection. I grew up on the original membrane keyboard 400, had an 800 and 1200XL, and collected a 130XE.

I ran into a similar thing running a dense LLM that doesn't use many MoE experts with vast knowledge. It ends up doing a lot of searching. In the end my local AI is as free as electricity but I pay for the searches.

The search providers I pay likely don't pay their sources. Some others (that I don't use) don't even have the original text only providing summaries. Some offer both.

Unless it's a personal project that you vibe coded, any production code that is the primary asset to a real company should have a person who understands what gets deployed. The best way I know how to do that is to co-design it with the AI. Reading code to match a spec can get the job done but the structure can be all kinds of horrible, and the more it gets iterated upon with features only makes it sprawl.

Two signals I use: do I ever push back and find places where the draft version was incorrect, incomplete, or otherwise not up to par and do I feel comfortable if something goes wrong with in production and I have to debug it at any hour of day or week.

And what's with needing a custom runner? Why not publish the model on Hugging Face so we can use it on any compatible with the model format?

For that kind of money I can run 2x AMD R9700 (on a server I already had) and get 150-250 tokens/sec running Qwen3.8-27B MXFP4. That speed makes all the difference as then the work is highly interactive design/development rather than semi-batch agent loops.

Thanks for confirming there's no technical reason for the omission!

Was DeepSeek Harness (DSH) intentionally omitted from the listed harnesses for any technical reason?

A lot of coding was actually plumbing filling in blanks between a framework and libraries with a few decisions in defining the data or process/flow. Not much lost there using AI where you still shape and review+understand the output.

I don't get the point. Running it on itself:

    Verdict for “https://openfatigue.ai/”:

    Formerly Open
    Openness score: 9%

    Was open until it got popular
    License changed on a Friday at 5pm
    Community forked it and called it OpenFormerly

Is this a joke site? First Pricing card

    Open

    $0/mo*

    Read the README
    Star the repo (required)
    Join the Discord (read-only)
    ~~Run the software~~ [strikeout]

    [Get Open]

It's the utility value I can't see. It makes sense for as much as you can put "into a laptop" and is the same zone as AMD Strix/Gorgon Halo. I'd much rather have the high compute in my server in the basement that I have remote use of rather than burning my battery with limited thermals.

And the Apple Max/Ultra does it better. The Mac Mini/Studio form factor makes way more sense. These laptops look like conspicuous consumption brand marketing devices. CUDA is worthless--AIs can port kernels.

The spec I haven't seen listed on any of these RTX Spark laptops is the memory bandwidth. Seeing as it's listed as having up to 128GB LPDDR5X, I can guess that it's about the same as an AMD Gorgon Halo's 192GB around 273 GB/s which is where I lose interest. Even the M5 Pro at 307GB/s can't compare to discrete GPUs.

Only when we get to M5 Max 614 GB/s and Ultra 1.2TB/s does it get interesting for me.

Who, what, when, where, how, am I?

    whoami
    uname -a; cat /etc/issue
    date
    pwd; curl http://ipv4.icanhazip.com
    df -h; free -h; # htop