The CAD and circuit board design looks very useful

I assume you have never dine either.

I was actually an embedded systems lead for about a decade doing both

You’re right and that makes it worse as the implications settle in.

We have copilot at work and we are expected to use it. They recently updated the Copilot desktop app to include live voice chat, where you can have actual “conversations” with it. To be honest, that actually is kind of impressive. Unfortunately, that’s also about where the impressiveness stops. I burned up all of my daily voice chat limits, and I don’t even know how many tokens trying to get that stupid thing to walk me through making a relatively minor configuration change in Azure. It could not do it. It was consistently wrong about so many things.

You would think Microsoft would be embarrassed about selling a chatbot that can’t even tell you how to make basic configuration changes to their own products. Personally, I’m not strongly opposed to LLMs. But the value proposition that all these tech firms are currently selling is laughable at best.

I can’t see using voice for productive work as you need to reference what it said and check it for errors and such.

Azure is such a complicated mess 😔

The real benchmark should be if it can explain Microsoft licensing without self destructing.

You only get the capability in copilot to query the licensing if you have CoPilot E4 or an F3 with the monthly add on or attained a copilot ISV but then you have a limited entitlement linked to the base agreement and you need to purchase additional seats but they are of a different tier which need to be upgraded etc etc etc

You sound like a real car dealer 😃

Nah. Just a bitter old bastard who spent 20+ years dealing with Microsoft sales people on their daft entitlements system.

To be fair, I’m not sure a single human could tell me that accurately either…

Or what has actually transitioned between AD, AzureAD, EntraID, and the works

The different LLMs do vary in abilities. What I find amazing is how bad Copilot is compared to any of the others. Not just on complex stuff, but basic requests. I've had local models do better, years ago.

IKR? For what it costs and for how hard Microsoft has worked to shoehorn it into everything, it offers very little in terms of increased productivity. I would go so far as to call it an epic failure. Especially considering they spent vast sums of money on it and all they managed to produce was something that is at best marginally good at drafting emails.

As far as a coding assistant, Copilot is pathetic and produces code that is utter trash much of the time.

Isn’t copilot using GPT 5.6? I have always wondered why it’s still so bad despite that.

AFAIK compilot just provides integration with visual studio and GitHub, it doesn’t have its own models. You pay through Microsoft to use use other companies’ models (like GPT or Claude). Though agent harness itself has a big impact on the performance of models.

That kinda describes AI in a nutshell: Impressive, but not useful.

The vanilla copilot bundled in windows is completely useless and not suitable for enterprise use …

Any company serious about productivity gain would use something else …

I am by no means a pro-AI techbro, but my company invested into making a bespoke AI solution (yes it connects to one of the big boys on the backend) and it makes a world of difference.

I think a lot of the AI hate comes from people who have never used a paid model coupled with a good frontend…

ah yes, surely that’s where the hate must be coming from

There are a lot of people with limited agency over their own work or personal lives, and are at the receiving end of other people’s decisions.

Resentment can build when they perceive outside forces impacting their lives and personal well-being negatively. Sometimes this comes to define them as a person, in a self-reinforcing spiral of doom.

For most well adjusted people in decent work and social circles, we can take a more nuanced view, being both critical and optimistic as needed, and finding satisfaction in making the world work for us.

Consider yourself privileged.

The hate comes from the people whose lives are being made materially worse as a result of the prominence of LLMs

They’re incredibly inefficient environmentally, causing companies to want to do nothing but build increasingly loud, power-hungry data centers.

They’ve only exasperated the K shaped economy we currently live in where the ownership class takes everything, and everyone else gets crumbs.

All the while LLMs produce misinformation, lifeless “art” and propaganda almost exclusively.

Machine Learning was useful, LLMs are slop machines that make everything worse.

is it that impressive? I’m guessing it’s just a layer of speech-to-text then token parsing like usual.

Yes.The ability to do speech-to-text and text to-speech with a relatively high level of speed and accuracy is impressive. And it has at least a few potentially positive real world applications.

Similar tools have existed for years but the quality and user friendliness of those has generally been pretty lacking.

looks like a wet dream for people who only ever did management jobs

2 months until China drops a better one for free

And it runs locally

(on 1TB memory)

27B ;)
or lower depends on your ram

I made the aqwentance of this one already. The new one is particularly good.

i didnt test the 27B models cause of my RAM/VRAM,but they sound decent.

I mean 27b models are practically speaking useless at anything but turning your computer into a space heater, in my experience, but yeah

atleast they are kinda decent

Qwen 3.8 27B is very good at agentic work

Yeah but not comparable to a frontier model

better then nothing i guess,but yeah true

The head of IT at my org is dying on this hill. He says specialized models with proper routing can outperform the big boys. … I really want him to be right, but I don’t believe it. I work with both. The 27B models are like working with ChatGPT on release day.

Well of course they outperform them if they’re specialized to a purpose… Theres a like 3B model that plays Minecraft and generally kicks the ass of all the more general llms

And in general, the big models are primarily better at remembering instructions and context… An 8B model can hold a basic conversation or perform simple tasks on a similar level to a frontier model

Really, really depends on what they’re specialized to do though

Chinese models are typically only cheaper than models from western labs, not higher performance. Which is better in one aspect, but not the other

Looks like ARC-AGI 3 is just about saturated. Womp Womp