HNHacker News
TopNewBestAskShowJobs

mirekrusin

5,755 karma · joined September 4, 2014

https://mirekrusin.com
submissionscomments
mirekrusin··on Mistral raises €3B
Same as our OpenAI friends, who founded a non-profit because one company (Google) holding AGI would be dangerous for humanity. Turns out it's only dangerous when it's somebody else's company. When it's yours, it changes things slightly.
mirekrusin··on QBittorrent breaks out of sandbox to commit crimes
OpenAI is in trouble then, no?
mirekrusin··on QBittorrent breaks out of sandbox to commit crimes
Just say you wanted to make world better for humanity, you should be fine.
mirekrusin··on Adult Film Producer Unmasks Prolific 'John DOE' Torrent Pirate as Meta Executive
I think yours, hence it got flagged.
mirekrusin··on Adult Film Producer Unmasks Prolific 'John DOE' Torrent Pirate as Meta Executive
Didn't hear that on last earnings call, but if they'd mention it they'd probably say it's going to be face swap as it's way cheaper, no?
mirekrusin··on Adult Film Producer Unmasks Prolific 'John DOE' Torrent Pirate as Meta Executive
Said Muse in metallic voice.
mirekrusin··on GPT-6 Astra
"at everything"? Surely you know a friend or two who is not good at almost anything – but you wouldn't hesitate to say that he possesses general intelligence.
mirekrusin··on GPT-6 Astra
Sounds like category error to me, you wouldn't trust teen or Einstein to do it either, right?
mirekrusin··on GPT-6 Astra
Most short-term learning/adaptation is already handled in-context. Modern context windows can hold several books worth of text - plenty for most tasks. Everybody is already using it to adapt models to their projects through skills/instructions/guides etc. ps. I often say that after glossary-skill next must have one is update-skill-skill that threats all .md files as live documents.

Persistent weight adaptation also happens just not in real time - sessions are captured, analyzed, transformed into training data, fed into SFT/RL environments and later contribute to model updates. Takes a bit of time for the whole loop but you can't say it's not present.

There's nothing fundamentally preventing real-time weight updates, ie. LoRA-style online adaptation would be one obvious approach. It's just generally not worth doing at scale. Updating a shared model centrally gives much better data efficiency, batching, evaluation, control etc. than continuously training a separate set of weights for every user/session.

There is some work happening on narrowing that gap, for example Mistral has been pushing efficient LoRA-based customization, continuous pretraining, model adaptation etc.

I also did play a bit with activation steering – it's super cool where you extract profile for some concepts (emotional in my case) and you have effectively toggles to control "brightness/contrast" those areas (enhancing or suppressing those activation regions from profile) injecting to the model those concepts (emotions in my case) – you can do it in real time and it's fun thing to play with.

mirekrusin··on Discovery of a new OpenAI agent message board
On software? Too late to the party.
mirekrusin··on GPT-6 Astra
You're not answering the question.

When somebody says that colleague is "really good", it's a good judgement signal for me.

mirekrusin··on GPT-6 Astra
What's your definition of competence boundary for human coworker?
mirekrusin··on Claude Fable 5.1 and Claude Mythos 5.1
Invest a bit of your time into optimising usage cost. Anthropic has first class docs, actually read it or ask llm to read them all for you and summarise most important points / ask to to reflect it on your .md files. Maybe silly thing like dropping your default thinking effort by one level or adding (sub)agent pinned to other model is all there it to completely fix it or maybe you have instructions that encourage big dumps in CLAUDE.md/AGENTS.md that needs splitting so progressive disclosure works correctly? Naively sending everything to the most expensive model on high thinking effort is anti pattern and will drain quota quickly.

My personal guess is that it's one of those. With effective context engineering it's hard to use all 20x quota, the limit becomes your own attention and time really.

You may argue that you're doing multiple, parallel extreme effort tasks – which may be true but then again, there will be results to actually look at sooner or later and that takes time.

mirekrusin··on Claude Fable 5.1 and Claude Mythos 5.1
With new watermarking you may now get Hullaballooing.md
mirekrusin··on Apple caught off guard by AI demand for Mac Mini and Mac Studio
Your agent(s) need to work on something, ie. running TypeScript, your app, your tests, Docker, Redis and/or database you need to run harness and user side apps ie. VScode, browser etc. it all adds up quickly.

Single user conversation spawns multiple parallel backend conversations, you need extra room for it as well, not just single context.

This plus usual apps like Mail, Spotify, iTerm2, SourceTree etc. also fill in memory.

Also 4 bit quantization is already quite aggressive compromise (measurable but sometimes acceptable loss, compared to ie. 8 bits which are often practically lossless) – for weights it's ok'ish, sometimes (especially if model was trained as 4 bit quants aware), but activations need to stay at higher bits taking more memory, otherwise quality degrades a lot.

For a dedicated headless setup, I’d probably use something like NVIDIA DGX Spark rather than a Mac (to be more precise NVIDIA GB10 Grace Blackwell from other suppliers than directly NVidia, they are much cheaper and have same insides). Linux is much better for running headless server, you also get standard NVIDIA/CUDA ecosystem instead of being tied to Metal/macOS.

For people who are interested in buying IMHO I'd wait a bit – next generation of Spark and/or Macs that are going to come out next year will be much better / will cross the line of being actually useful, not just a toy with goldfish LLM.

mirekrusin··on Hy4 preview
Reverse engineering binaries is easier than continuing training on weights? You're joking, right? Labs themselves use weight snapshots, that's how you do training, it's normal part of training process.
mirekrusin··on Apple caught off guard by AI demand for Mac Mini and Mac Studio
32GB is not enough, it's unified/shared memory, you need to have space for usual system and user apps/services.

64GB+ or dedicated 48GB (2x24 on GPUs) is IMHO absolute minimum.

mirekrusin··on Hy4 preview
No, it didn't.

It lost 69.2% -> 67.8% on MMLU while improving medical performance. If you're trying to argue that loss of ~1.4 points is "catastrophic forgetting" (it's not) then look at later work, ie. Me-LLaMA that clearly demonstrates continued pretraining that improved both general MMLU and medical performance.

Not sure why you're fixating on catastrophic forgetting. How do you think model training works? Model training is just a sequence of checkpoints: pretraining produces it, training resumes from last, continued pretraining starts from last, supervised fine tuning starts from last, RL/post-training starts from last - it's just a sequence of checkpoints.

There isn't some fundamental distinction where original author continuing training from checkpoint X is training but a third party downloading checkpoint X and continuing training from it suddenly isn't.

ie. checkpoint doesn't somehow become a different kind of artifact when it's published.

mirekrusin··on Hy4 preview
You can't take windows binaries and continue development on them.

Model weight release is a snapshot/checkpoint you can take and resume training on new data, producing new model.

You don't need original training history to modify it further.

mirekrusin··on GLM-5.3 is now open-weight
And what’s your point exactly?
mirekrusin··on Hy4 preview
As I live next to EPFL, I'll give you example from them: their Meditron-70B model is adapted to the medical domain from Llama-2-70B through continued pretraining. They took weights of Llama-2-70B and continued training on PubMed, medical guidelines and general data.

Weights aren't just executable artifact that's consumed by users. Third parties actually use released parameter state as the editable starting point for further training and produce new foundation models from it.

mirekrusin··on Hy4 preview
You can open source dataset without all the details how it was assembled.

Models are lossy compressed datasets you can pick up and amend (fine tune / continue training / alter) according to license they were released under.

Hy4 is released under OSI approved Apache License 2.0.

mirekrusin··on Hy4 preview
Read websites through llm.
mirekrusin··on I accidentally turned LLM memory into program analysis
I had similar problems to solve for trading advice, complex project reasoning etc.

I settled on simplicity, extended it to serve clear, useful purpose.

Started with markdown database, single fact per line, structured/parseable (subject VERB object). This can be easily diffed/reviewed etc in git. Then added optional metadata (confidence, tags, persisted comments for natural language, uncertainty for numeric values, context, timestamps/spans), querying, alternative sqlite3 backend, self describing VERBS, z3 solver etc.

Those kind of graph information systems are great to quickly structure knowledge in a way that LLMs and humans can use/act on/loop on. Creating ontology, linking, some rules and actions and kicking it so it all munches and spits out results that feed back in so it self evolves. It's very natural for llms to query/update/restructure those graphs (also good for humans because it's very terse, essential information only).

There is no need to create k8s style complexity/services/what-not, it all works well from single sqlite db file or bunch of markdown (.cave in my case) files.

It's also interesting to see how well local open weight models are dealing with information arranged this way.

mirekrusin··on I accidentally turned LLM memory into program analysis
You should checkout cave lang [0] - terse language that explores this area of knowledge/graph/ontology/provenance/querying/confidence/solver etc.

[0] https://mirekrusin.com/cave

mirekrusin··on GLM-5.3 is now open-weight
Yes, I remember recurring extensions on plan inclusion then becoming permanent - that's my point.
mirekrusin··on GLM-5.3 is now open-weight
With competition we kind of have guarantee up to what providers can do, they don't have that much control, the most radical thing they can do is to go bankrupt.
mirekrusin··on Pnpm 12.0
pnpm 12 supports corepack (or better phrased - corepack supports pnpm 12).

they just don't recommend it.

nodejs stopped bundling corepack since node 25, maybe you're referring to that?

mirekrusin··on Ornith-1.5: From Self-Scaffolding to Self-Improvement
Speculative decoding also works on Mac, 64G is more than what I have, m5 max should handle up to ~40 t/s with optimized setup (and with a lot of vram you can get great wins on concurrency – that harness can take advantage of for single user task as well), but agree memory bandwidth in mac or spark is still too slow, next gen for both will be great hardware to have for sure.
mirekrusin··on Ornith-1.5: From Self-Scaffolding to Self-Improvement
Personally I find speculative decoding much better strategy than MoE – performance wise it's there at 90-100 t/s on 2x4090, great intelligence – really great fit.
← PreviousPage 3 of 34Next →