novaRom

The debt lasts longer than GPUs remain profitable, right? GPUs loose value most quicker than most investors believe. In best case they last like 2-3 years under heavy stress, so I think about all those data centers and what will it mean to replace those cards burned down - you cannot even replace one with newer one - all 8 GPUs should be the same type. Thinking about future costs with all the debt.

Yes, anything, not just software. I wonder if monopolies will somehow start falling or if they will try to defend themselves, are we on a path of continuous decentralization?

I wonder how many chat logs they did use in model's training data? If model consumes human’s reasoning (the chat logs) and then presents results as its own, it's a kind of parasite right?

Harvesting collective output of human consciousness to build AI that then replaces the very people who provided the input.

Mac's German keyboard was always pain for writing code. This is less relevant today because most of the input has become natural language so no need for brackets and other non alphanumeric symbols (which are terribly painful key combinations on Mac German keyboards).

I did use Mistral Web Chat a lot in the past years, but somehow went away because their UI become messy and default model has become a bit stupid in comparison with alternatives. Yesterday I tried it and compared with all other alternatives: Gemini, Mistral, Grok, Claude, ChatGPT (mobile apps, no subscriptions, default models), same prompt (a problem required sophisticated thinking and web searches), and surprisingly it worked better than any other free app. Closest was Grok which felt faster, but Mistral provided best solution.

If AI tech is so critical, then we have to do it completely open: any frontier AI company should become completely transparent - full public access to training data and recipes.

Aggressive marketing (including daily posts on HN). Many people simply don't know about alternatives, there are many people who never heard about pi and opencode and live comfortably in claude-codex bubbles.

> Mathematics has very little to do with the real world

But it's a construction based on very basic properties of the real world, at least on those properties we accept and perceive as very general and we use them as building blocks.

It will not just affect big software monopolies, but also hardware designer companies, cloud and streaming services. What the point to buy or pay subscriptions if you can reimplement it in a matter of days if not hours? A lot of business opportunities for small (even solo) enterprises. That's the power of AI revolution.

even with 256 bytes: https://dosasm.com

When one player tries to monopolize AI tech and market by all means, others are those who don't want to be hooked to a foreign will in the future. I think this is the reason in doing open research, and it's likely positive for all of us.

Yes, it happens quite often recently. I think, their approach is "dominate and tolerate" in a controllable way, so they know quite a bit about who is actually using NewPipe, yt-dlp, etc. and they tolerate it in order to have a kind of leverage to preserve their monopoly. Then if number of out-of-official-app views becomes like above certain point, they aggressively block with "bot detected" message.

Compaction is painful if you run just one local LLM, the best way to avoid it is to keep context as small as possible.

One trick I find useful is to have one model with two KV caches running and while first cache has produced tokens, second cache immediately summarizes them during input tokens are being generated (tools time), then harness switches to the second KV cache which takes newly produced input tokens while KV in first cache is getting replaced with compacted summary tokens. This is a kind of ping pong, so we trade more space for less time. Still experimenting but it looks it works, and nice bonus it improves GPU utilization. Btw I have my own harness and model serving code, but it can be easily implemented in any other harness and model server.

it said me it is llama 4 1.5B

Not just you. Hassabis in a recent interview looked a bit nervous when talking about AGI is rather sooner than many expected.

But so is German industry - every day in the news either insolvency or workforce reduction and moving production to Eastern Europe - main reason often stated "expensive Energy". I wonder what are the stats for industrial production in 2025 or total energy consumption in Germany?

> It clearly memorized the code

This code could appear many times in training data; but I wonder what decoding params did you actually used, because that effect appears likely with close-to-greedy way. Also, if it is a coding assistant model, chances are this skill was increased by additional sampling from that domain during training? Or maybe RL has teached model to provide verbatim answers for certain questions? So it was rewarded if that kind of answers is really helpful (I guess it was?).

Big shift will happen after their Q2, 20% of eligible shares release; directed-share (friends & family) participants may sell for the first time.

Q2 call is expected in a month or so? But they also state Aug 21 is 37% of shares will go public.

https://spcx.capital/spacex-stock-lockup-dates

Gemma 4 12B QAT is amazing - agents run very fast, and it's really very smart, at least in my agent's harness domain which is GNU software development - on par with frontiers like GPT Sol, DeepSeek, or Claude - Why to buy those expensive tokens if a local tiny model performs so well?

>> At the moment, we are locked in an extremely intense, multilayered commercial and geopolitical race.

* Creating (and replicating) any software which can run in isolation locally is completely solved - we already see initial tries to create completely new operating systems, messengers, tools and platforms

* Replicating large Web-services like Netflix, Google Search, YouTube, or TikTok is more difficult but not impossible - national regulations may cause major and quick shift here. Tech is there, initiative is missing yet. Next 2-4 years we'll see new Amazons, Netflixes, and Spotifies.

* Remaining "hard" blocks are due to hardware production monopolies, but it's a matter of months or few years: we will see more and more new GPU types with new software frameworks, completely from scratch, covering all stages, incl. full physical layout - tapeout ready; new CPU and DRAM architectures; EDA tools - HPC-ready, GPU-aware

Yes, Singularity is near ...

Have you tried most powerful recent models? You may even don't understand at all "how" - what you get is end result you asked for, verified and tested. I am pretty sure if I can vibe code quite non trivial solution in just say 4-8 hours today, then why will it be impossible say in a year or two to vibe code something at the level of most sophisticated apps and services today? Pretty reasonable expectation taking into account what was possible just a year ago and what you can make today without even understanding "how".

Exactly! I just vibe coded (with GPT Sol and Claude whatever-number) my own agent, it's trivial to add now any feature I want - simply ask more powerful model to do it for you. I am happy with end result, however it looks indeed these tools are trained to increase token count - they do quite stupid token-spending steps while making code, but the code itself is also a bit weird - it's like they intentionally do code which is hard to modify on your own without using exactly those authoring models. Interestingly, when I am using DeepSeek with OpenCode, I don't see that - it understands my intent well enough and overall code quality is not bad. I recently switched to local Gemma 4, and I often switch (in opencode) to just that less powerful model, because it understands my intent and has enough skills to provide good quality solution although it's rather for small size projects, and for not coding from scratch, but it's also free and private. It feels slower than any big cloud model, so my model switching is probably most quickest path to robust end result :)

It's probably a matter of short time until it's possible to disassemble any sophisticated software, rewrite it entirely with better features and usability, generate all needed artifacts, port to any platform. The only moat left is probably remote massive data storage. So if you want to replicate YouTube or TikTok, it's not impossible, but requires a lot more hardware assets than say anything that runs entirely locally (like operating systems or most video games).

I haven't tried Codex yet, but I for my tasks GPT-5.5 may correctly point to a proper direction but its code feels a bit weird. Opus 4.8 is way better in coding, and actually it's the only one who could catch very very sophisticated bug in a large codebase (I tried different models including GPT-5.5 and DeepSeek). Interestingly Gemma 4 under opencode running locally performs not bad at all, it's far yet from DeepSeek level, but it manages to understand tools quite well, and code quality is pretty good. So, for simple coding projects I can say local models already won. It's amazing how smart open models of desktop size have become today. I mean it's quite plausible to manage small codebase today relying on only open tools and local models, you don't need any subscription to produce high quality code, but yes I assume you already experienced and know what you're doing :)