[–] 8 points 4 months ago (1 child)

I have been running renovate on my yaml files, been really happy with it. Even got notifications set up finally so it tells me what gets upgraded to what. Although for me it is not a constant container but a Cron job that runs in my k3s cluster (since I'm using k3s anyway)

  • source
  • [–] [S] 1 point 4 months ago

    Yet this is my problem, A tool that "waste" tokens, is slow (with some tasks taking 30~40 minutes) and generating a result, while a model running locally can do the 70% in 5min. It's a huge difference.

    And don't forget, at the end of the day, I would still need to look into the code and fix some stuff myself.

    And if compared with Deepseek (which had fit best on my use case), it's even faster like 1~3 minutes, to get on the 90% of the result, on a fraction of the cost. This is the points that I'm focused, and a enterprise should either understand that the market have changed, or pay the price. And looks like we will pay the price.

  • source
  • parent
  • context
  • [–] 2 points 4 months ago (1 child)

    Claude is geared more towards enterprise companies, anyway. I think you went over pretty succinctly as to why: they can afford the extra resource burn.

    pretty much the way I use AI rght now is similar to how you see the near future of your use.

    • AI does first draft
    • I review draft, conclude 'this is shit'
    • redo most of it

    its pretty useful for getting through writers-block type scenarios sometimes because of that, but past the initial stages of the project the usefulness declines dramatically.

    I'm still not seeing any real ways of removing the human in the loop, especially for the review process.

  • source
  • parent
  • context
  • [–] [S] 2 points 4 months ago (2 children)

    Thanks for the reading, the original post was somewhat unhinged. It was a very long debugging session for the alpha thing, the AI tooling weren't helping. And I was losing my mind over my morning coffee. But I fixed the little details.

    I think the reality of the AI tooling for coding fall in the "middle", it isn't the 10x or 100x described by people, it's closer to a 10%~15% performance increase in my humble view. With some better harness (like the PI that I discovered) I think this can be sightly better, but far from the 50%. I would say that if the increase in productivity decays over time, the amount of code generated I had to refactor and clean for most of the week.

    But for sure the Claude was a huge disappointment, very bad quality for it's price, and super slow for the needs of a "start up mindset", let's put in this way.

  • source
  • parent
  • context
  • [–] 2 points 4 months ago (3 children)

    thanks for your writeup. I didn't read the original but this one was fun to read through.

    this looks like an awesome project, and I appreciate how you went over the ways AI helped and hindered your project at the end. its really hard to get honest feedback about AI from devs IRL right now, everyone keeps saying they have silver bullets in their workflows and I feel like the crazy one for not being able to figure it out.

    looking forward to giving AC a shot!

  • source
  • [–] [S] 2 points 4 months ago

    I started this project in January 2010, and I have been maintaining it ever since. Only recently did I use Claude Code with Opus to make some improvements to the layout (mainly the third column and the new scrolling system) and the README. I'll continue to use Claude Code to help me with UI and also maintain the translations.

  • source
  • parent
  • context
  • [–] 1 point 4 months ago* (last edited 4 months ago) (2 children)

    Curious why do you swap between Qwen and E4B. On my hardware they perform with similar tps. Qwen 3.6 35B spits out 80-100tps on AMD 9700 and E4B gives me about the same tps.

  • source
  • parent
  • context
  • [–] 5 points 4 months ago (1 child)

    I know you said not clean, but Claude and other LLMs with this ability can and do still hallucinate even when researching. You can have it give you direct source links (not just the attributes that it gives by default) with explicit instructions to quote the exact finding, word for word, and sometimes it will suddenly tell you that it was made up and not actually in the attributed page. Not every time, but it's something to be very careful with.

    Google's own LLM struggles with this issue as well.

  • source
  • parent
  • context
  • [–] 2 points 4 months ago

    Thank you so so much for pointing out ROCmFP4. I have been tinkering with my RDNA 3 framework on llama. I was struggling with ROCm llama.cpp and have been using vulcan in the meantime. I know there’s some issues on the llama.cpp github to try and fix my issue (UMA stuff), but haven’t come across this specific project. Gonna try it out

  • source
  • parent
  • context
  • [–] 2 points 4 months ago

    What my Qwen 3.6 35B setup spat out to this prompt. Used 20-something sources:

    Based on my research, here's a comprehensive overview of the current state of nuclear fusion research:

    Current Major Breakthroughs

    The field has seen remarkable advances recently:

    • National Ignition Facility (NIF) achieved fusion ignition for the seventh time in February 2025, setting a new energy gain record of 2.44x. In April 2025, they produced a record fusion yield of 8.6 MJ - that's more than four times the laser energy put in (2.08 MJ).
    • Global investment has surged: Fusion startups have raised over $13 billion to date, with 17 companies having raised $100M+ each.

    Biggest Hurdles & Challenges

    1. Materials Science

    The fusion environment is extremely hostile. Neutrons from fusion reactions will bombard the reactor's first wall and vacuum vessel with more than 10 displacements per atom (dpa) per year, causing:

    • Severe material degradation
    • Tritium retention issues
    • Erosion of plasma-facing components This requires entirely new materials that can withstand decades of operation.

    2. Engineering Scale-Up

    Moving from physics experiments to power plants presents enormous engineering challenges:

    • For inertial confinement (NIF), you'd need to ignite pellets many times per second—described as a "dizzying" challenge
    • Maintaining stable plasma at 150+ million°C for continuous operation (not just short pulses)
    • Heat extraction systems that can convert fusion heat into electricity efficiently

    3. Plasma Physics

    Understanding and controlling plasma turbulence, instabilities, and confinement remains an ongoing science challenge. While tokamaks are proven, consistent, optimal performance is difficult.

    4. Tritium Breeding

    Fusion reactors need tritium (which is rare), so they must breed their own in the reactor blanket—a technology that hasn't been demonstrated at commercial scale yet.

    5. Cost & Complexity

    Current fusion devices cost billions to build and operate. Economic competitiveness with other energy sources (and fission) remains uncertain.


    Next Steps & Timeline

    Public Research:

    • ITER (the international tokamak project): Construction is progressing but the first plasma has been pushed back to 2034, a 9-year delay from earlier plans. Full deuterium-tritium operations are targeted for 2036+
    • NIF: Continuing experiments with higher yields, working toward more efficient ignition schemes

    Near-term Goals (2025-2030):

    • SPARC tokamak (Commonwealth Fusion Systems) aims to demonstrate net energy by mid-2020s using new HTS magnets
    • Continued NIF physics campaigns to understand ignition optimization
    • DEMO reactor design work for next-generation demonstrations

    Commercial Timeline:

    The Fusion Industry Association states that the majority of fusion companies anticipate commercial power by the mid-2030s. This seems aggressive but reflects accelerated progress and massive investment.


    Private Sector Research: How Promising?

    The private sector is arguably making the most exciting progress:

    Major Players:

    • Commonwealth Fusion Systems (CFS): Developing SPARC using high-temperature superconducting magnets to create a compact, cheaper tokamak. They've already built and tested their HTS magnet technology at MIT's Alcator C-Mod and are raising capital for the full SPARC device.

    • Helion Energy: Pursuing a unique pulsed magnetic fusion approach (not a tokamak), planning a pilot plant by mid-decade with backing from Amazon's Jeff Bezos and Bill Gates.

    • Tae Technologies: Founded by John Nuckolls (NIF physicist) and others, using an inertial electrostatic confinement concept.

    • 17 startups have each raised $100M+, collectively attracting over $13B in total investment.

    Strengths of Private Research:

    • Speed: Smaller, more focused teams can iterate faster than billion-dollar international projects like ITER
    • Innovation: Willingness to pursue alternative approaches beyond tokamaks (stellarators, field-reversed configurations, laser-based ICF)
    • Funding momentum: Strong VC interest and tech billionaire backing
    • Regulatory progress: The ADVANCE Act (passed 2024 in the US) creates a more favorable regulatory framework for fusion

    Concerns:

    • Overhyped timelines ("commercial by mid-2030s" may be optimistic given historical nuclear project delays)
    • Many companies will likely fail; it's an unproven industry
    • The "10-year away" phenomenon has plagued fusion for 60 years—can private sector avoid this?

    Bottom Line

    Nuclear fusion is in the most promising phase of its history. We've definitively proven scientific net energy (NIF), new technologies like HTS magnets are enabling smaller reactors, and private investment is unprecedented. However, the gap between "net energy in a lab experiment" and "reliable, economic electricity on the grid" remains massive—possibly 15-20 years still.

    The consensus seems to be: fusion might reach early commercial deployment by the mid-2030s from private companies, but widespread adoption is likely decades away. Whether that's good enough depends on your patience and how you weigh potential reward against current climate urgency.

  • source
  • parent
  • context
  • [–] 3 points 4 months ago

    Open WebUI + SearXNG + llama.cpp + Qwen 3.6 35B + 16-32GB GPU. Gives you 256K context and runs with 80-100tps on 3090. If you have less VRAM like 16GB it'll be slower but still probably tens of tps on anything recent. I run it on AMD Pro 9700 which is about as fast as 3090.

  • source
  • [–] 2 points 4 months ago

    AMD Strix is an APU, optimized for AI. It is the cheapest option I am aware of to run bigger models at home. 2k for 56GB VRAM, and less den 300W total power Budget.

    One could run smaller models. But for the context sizes required for research work, that is nearly impossible.

    Also, external services, like openrouter, can be used to use models hosted in the cloud.

    But for self hosted, you need something that can run models with at least 15GB of VRAM + Context. For comparison. Our highly quantized model uses 20GB of vram. For our 4 slots we need another 20GB on top of it (around 5GB for 254k tokens), making it 40GB.

  • source
  • parent
  • context
  • view more: ‹ prev next ›