smcleod

13 ms·

241.
by smcleod 1y ago
Aren't cohere's models pretty dated now? They don't even show up on leaderboards (synthetic or real) these days. What about GLM 4.5, Qwen 3 235b 2507 or even just Qwen 3 32b 2507 etc...
242.
by smcleod 1y ago
Interesting to see its source is missing from the GitHub repo. A pivot to closed source perhaps?
243.
by smcleod 1y ago
> Hello! I'm GLM-4.5, a large language model developed by Zhipu AI (智谱AI). I'm based on the GLM (General Language Model) architecture, which is a transformer-based neural network design.
244.
by smcleod 1y ago
Seems to be down? Lots of console errors with the likes of "Content-Security-Policy: The page’s settings blocked an inline style (style-src-elem) from being applied because it violates the following directive: “style-src 'self&#x2
245.
by smcleod 1y ago
I'd much rather run a (native) app than have yet another browser tab. What I don't want is bloated Electron apps.
246.
by smcleod 1y ago
Have a watch of https://youtu.be/vW30o4U9BFE
247.
by smcleod 1y ago
Have a watch of https://youtu.be/vW30o4U9BFE
248.
by smcleod 1y ago
My thinking is that across software development as an industry we have significant subscription fatigue. Obviously for hosted services there are ongoing costs involved - and regardless businesses should be paying people for their work, howe
249.
by smcleod 1y ago
Do you plan on releasing the self hosted version free (or better yet - open source) for non-commercial, personal use?
250.
by smcleod 1y ago
Within the GGUF (and some other formats) you'll see each layer gets its own quantisation, for example embeddings layers are usually more sensitive to quantisation and as such are often kept at Q8 or FP16. If you run GGUF-dump or click
251.
by smcleod 1y ago
No they don't. GGML is non-uniform. Each layer gets its own level of quantisation.
252.
by smcleod 1y ago
There is no reason to run models for inference at static fp16, modern quantisation formats dynamically assign precision to the layers that need them, an average of 6bpw is practical imperceptible from full precision, 8bpw if you really want
253.
by smcleod 1y ago
Claude 4 Opus when provided with the full paper, blog post and asked to perform a critical review: The methodology contains several fundamental flaws that likely explain this anomalous result. Most critically, the study examines a highly sp
254.
by smcleod 1y ago
The framework desktop isn't really that compelling for work with LLMs, it's memory bandwidth is very low compared to GPUs and Apple Silicon Max/Ultra chips - you'd really notice how slow LLMs are on it to the point of fr
255.
by smcleod 1y ago
I've taken a very similar approach on a number of monolith decomposition projects over the past year and a bit. They've mostly been big, old, tightly coupled Java - but it's worked very well.
256.
by smcleod 1y ago
Python as a language is nice. Python's version and package management is nothing short of a nightmare.
257.
by smcleod 1y ago
It's not that simple: https://github.com/aws/amazon-q-developer-cli/issues/844
258.
by smcleod 1y ago
I'm the one that logged this: https://github.com/aws/amazon-q-developer-cli/issues/844 Amazon really have dropped the ball with fixing it.
259.
by smcleod 1y ago
My 2c-TLDR; after doing some battle testing over the few weeks (so keep in mind some of this was on pre-release versions): The good: It's great to see they've baked in the concept of setup -> plan -> act into the tool with
260.
by smcleod 1y ago
Q CLI has so many issues though, injects so much junk into you shell profiles it can slow down your terminal invocations by seconds, and it doesn't support standard streamableHttp / SSE MCP servers.
261.
by smcleod 1y ago
Oh! I'm sorry, I must have misread!
262.
by smcleod 1y ago
That's certainly not the case in the majority of Europe, Australia, New Zealand, East Africa, West Africa, South Africa, PNG, Fiji, Philippines, South Asia and South East Asia. Really it's mostly just America and Canada that use A
263.
by smcleod 1y ago
"The only whole world uses American English" ... I think you might be living in a bubble. It sticks out like a sore thumb.
264.
by smcleod 1y ago
Well, only Americans use it. There's no point in arguing about it your version of a language is better or worse but for the rest of the world it's incredibly annoying having to correct Zs with Ss when using LLMs or American only s
265.
by smcleod 1y ago
How well does it handle standard international English? So many of the tools I've seen seem to only support American English.
266.
by smcleod 1y ago
Good on you for taking action on this kind of feedback!
267.
by smcleod 1y ago
The $20/month cursor sub is heavily limited though, for basic casual usage that's fine but you VERY soon run into its limits when working at any speed.
268.
by smcleod 1y ago
Honestly there's no value that windsurf, cursor and all the other VSCode forks provide that couldn't be provided as an extension and even then - none of them perform as well for agentic coding as Cline / Roo Code (debates abo
269.
by smcleod 1y ago
FYI I see you have SmolLM2, this was replaced with SmolLM 3 this week! Would be great to have a few larger models to choose from too, Qwen 3 4b, 8b etc
270.
by smcleod 1y ago
That's about right for copilots. It's much higher for agentic coding.

More ›