mft_

10 ms·

151.
by mft_ 4mo ago
This is always how I've imagined small/consumer-hardware models going in time. If I only ever code in Python, give me a model that does just that (plus some general CS, algorithms, structure, etc.) and does it super-fast and well
152.
by mft_ 4mo ago
Eh, humans give names to and/or anthropomorphise lots of things. My partner names all of her cars and bikes; I don't. Isn't it more rational to feel some sort of connection and anthropomorphise a tool with which you can at
153.
by mft_ 4mo ago
The lowest spec M4 Mac Mini on apple.com is $799 today. The next generation Mini will likely be more expensive due to memory pricing, and as the Steam Machine already includes current higher memory pricing, that would be a fairer compariso
154.
by mft_ 4mo ago
> And turns out Virgina Tech does a bunch of helmet impact testing and maintains a ranking list https://www.helmet.beam.vt.edu/ Thanks for sharing. Interesting to see my Giro (with MIPS) has... 3 stars. Hmmm.
155.
by mft_ 4mo ago
In your take, why are they producing and giving away such good local models? Mindshare? Promotion?
156.
by mft_ 4mo ago
Wow… so even hidden behind the apparently good, user-friendly change… is underhand toxicity?
157.
by mft_ 4mo ago
I disagree - I think it might be more about predictable-hassle-avoidance. The media landscape is all about creating sensation, trying to find an edge from something a politician said or a corporation did to generate a headline and some clic
158.
by mft_ 4mo ago
Often, the experience is of initially trying to avoid shaving the yak, but finding all of the available alternatives are unusable or unsatisfactory for some reason or other. So yes, you shave the yak, but at least the clippings are precisel
159.
by mft_ 4mo ago
I’m using Hermes at the moment - it comes with lots of tools already baked in for the agent to use - for example web and browser access just worked, rather than having to mess around loads with config scripts and plugins. I’ve also tried Op
160.
by mft_ 4mo ago
There are likely huge unmeasured inefficiencies built in to many companies. You hear anecdotes of people whose job for years has been to perform a particular task, which could be either automated or performed more efficiently (by orders of
161.
by mft_ 4mo ago
Something commercial? Mistral is the only significant European LLM I'm aware of, and it's not frontier level. Then a dearth of EU-owned hyperscalers, AI-focussed data centers, GPU/CPU/memory manufacturers, even cheap e
162.
by mft_ 4mo ago
Where would they go? 1) It’s safe to assume the US would do its best to prevent it, and even if Anthropic was successful in exfiltrating their data, code, models, and people, I’d imagine the US would immediately block all US companies from
163.
by mft_ 4mo ago
I have a Mac with 4TB of storage but it’s still annoying when every new AI app I try installs its own virtual environment with a fresh copy of Python, PyTorch, other duplicate libraries, and then models on top of that.
164.
by mft_ 4mo ago
Thanks, I'll give it a go. (I generally find standard 27B too slow to enjoy using, whereas 35B-A3B is pretty snappy.)
165.
by mft_ 4mo ago
I found a marginal downside to Qwen3.6-35B-A3B-MTP vs. the non-MTP equivalent on an M1 Max. I’ll maybe experiment with settings further though.
166.
by mft_ 4mo ago
Tangent: did the MTP help you at all? I’ve tested that model back to back on my M1 Max MBP and the MTP version was actually marginally worse. I wonder if I didn’t use the right settings, although I tried several based on the obvious sources
167.
by mft_ 4mo ago
Can you quantify what it is you don’t like? Like, to my eyes ‘original’ is fine - and it’s very similar to ‘QT’ expect with rounded corners and brighter colours.
168.
by mft_ 4mo ago
As you note, I wonder to what extent this is a harness issue? I've been experimenting with different harnesses for local models, and with (IIRC) Hermes and Qwen3.6-35B-A3B I was amazed the lengths it went to (writing test code, opening
169.
by mft_ 4mo ago
Yes; but they don't (AFAIK) have a track record or pre-existing facilities for car design or manufacturing, per the comment I was replying to.
170.
by mft_ 4mo ago
I'm watching Slate [0] with interest; they aren't following a path according to your assumptions. [0] https://www.slate.auto/en
171.
by mft_ 4mo ago
Move to Europe, friend - a weight will be lifted.
172.
by mft_ 4mo ago
Speaking as a Brit, I couldn’t disagree more. I have no trouble understanding a wide variety of Europeans in a corporate environment, but sometimes struggle to even understand the basics of what Americans are trying to communicate, let alon
173.
by mft_ 4mo ago
I always remember a posed photo that one of my old bosses had on her desk. It was of her and her daughter; she was giving a big attractive (to my eyes faked for the camera) smile, and her daughter looked miserable. I appreciated the uninten
174.
by mft_ 4mo ago
Yes, before the enshittification kicks in. 3D printers are mostly still repairable and far more reliable and usable than a few years ago when the majority of the hobby wasn’t making stuff, it was tuning the printer to work reliably. Bambu a
175.
by mft_ 4mo ago
I was discussing with a colleague recently that this is becoming more and more of a problem: people delivering significant (sized) pieces of work which are obviously 99% written by LLM.
176.
by mft_ 4mo ago
Would you either share the code, or describe the approach? I’ve asked LLMs how to do something similar a few times recently, and received several different potential approaches. I’m still no clearer which route to take!
177.
by mft_ 4mo ago
> and either build a payment processor Risk, time, complexity, mismatched skill-sets... "getting a new thing built" didn't work so well with (for example) Fujitsu and the Post Office, or the billions spent to little or no
178.
by mft_ 4mo ago
Is this [0] saying that unsloth's versions of Google's QAT models are better than Google's own QAT models? Or am I not understanding it correctly? [0] https://unsloth.ai/docs/models/gemma-4/qat
179.
by mft_ 4mo ago
To expand on another answer, it’s all about the harness. Different harnesses (Claude Code or Cowork, Hermes, OpenCode, Pi, etc.) offer different default tools, system prompts, and ultimately approaches. (IME the corporate CoPilot app is ter
180.
by mft_ 4mo ago
To me, with experience of numerous organisations of different sizes outside tech, it is pretty surprising. (I'm not arguing that this is right but) the typical progression of an organisation as it scales is to move away from the '

More ›