It is not llm specific. A large swathe of it isn’t that much Microsoft specific either.
And it is a developer feature hidden from end users.
e.g. - In your ollama example, does the developer ask end users to install ollama? Does the dev redistribute ollama and keep it updated?
The ONNX format is pretty much a boring de-facto standard for ML model exchange. It is under the linux foundation.
The ONNX Runtime is a microsoft thing, but it is an MIT licensed runtime for cross language use and cross OS/HW platform deployment of ML models in the ONNX format.
That bit needs to support everything because Microsoft itself ships software on everything.(Mac/linux/iOS/Android/Windows.
The primary value claims for Windows ML (for a developer using it)—
This eliminates the need to:
Bundle execution providers for specific hardware vendors
Create separate app builds for different execution providers
Handle execution provider updates manually.
Since ‘EP’ is ultra-super-techno-jargon:
Here is what GPT-5 provides:
Intensional (what an EP is)
In ONNX Runtime, an Execution Provider (EP) is a pluggable backend that advertises which ops/kernels it can run and supplies the optimized implementations, memory allocators, and (optionally) graph rewrites for a specific target (CPU, CUDA/TensorRT, Core ML, OpenVINO, etc.). ONNX Runtime then partitions your model graph and assigns each partition to the highest-priority EP that claims it; anything unsupported falls back (by default) to the CPU EP.
Extensional (how you use them)
• You pick/priority-order EPs per session; ORT maps graph pieces accordingly and falls back as needed.
• Each EP has its own options (e.g., TensorRT workspace size, OpenVINO device string, QNN context cache).
• Common EPs: CPU, CUDA, TensorRT (NVIDIA), DirectML (Windows), Core ML (Apple), NNAPI (Android), OpenVINO (Intel), ROCm (AMD), QNN (Qualcomm).
I zoomed out. It looks like this:
"Usage has absolutely declined from peak switching periods where inevitibly some users won't stick around, but that's to be expected"
That just isn't a "sharp decline" no matter how much you seem to want to repeat those words.
I agree that it’s not a sharp decline but zooming out, what I see is absolutely no organic growth at all in the past couple of years. All the increases have been sharp spikes that immediately fall off dramatically, followed by longer, slower periods of decline. It looks like nobody is switching to Bluesky except in a handful of viral events, during which a tonne of people try it out but don’t keep using it. There’s only one upward slope on these graphs, and that stopped in late 2023 – about the time Threads went fully global. These look like very unhealthy stats.
Those are Jaz’s daily unique action counts (flows) from the Bluesky firehose; they’re anchored to the Nov ’24 spike, so the ‘decline’ is post-surge reversion. Meanwhile the user stock kept rising (~39M).
A presidential election spike is the baseline for tracking growth in a social media platform??
What’s “user stock”? Is that the number of registered accounts? Isn’t it basically impossible for that to do anything but go up? It’s the number of people actively using the network that’s the important figure, not the total number of people who ever used it.
Are these the figures you are reporting?
> We made a new Bluesky stats page to see how the platform is growing. Unfortunately it is currently shrinking.
> Last week the total number of users registered hit 36M, but actually only 13M of those showed any activity in the last 90 days.
Right, so the basic story you're telling investors there is that between August and September of 2024 they experienced a sharp spike, and then basically they stayed that way for over a year. That's not a dying platform, but it's not a growth story you take to investors either.
How does it compare to other social networks like Twitter? Can't compare because they don't offer granular data this detailed? That tells you something.
That doesn't matter! In fact, Twitter doing worse while Bluesky usage is dropping probably makes them significantly less investable.
I'm not rooting for them to fail. I use Bluesky. I find Twitter's ownership odious and the platform significantly worse than it was 4 years ago.
But if we're talking about scientific communicators talking about where the future of scientific communication is going to happen, it is relevant whether Bluesky has a long-term future. There's another non-Twitter social network that doesn't operate under this funding pressure!
Personas are a great tool. IMO - By the time you arrived these had transformed into bad shorthand. (I say this having been in Devdiv through those years.)
Elvis is not a persona - it is an inside baseball argument to management. It suffered a form of Goodhart’s law … it is a useful tool so people phrase their arguments in that form to win a biz fight and then the tool degrades.
Alan Cooper, who created VB advocated personas. When used well they are great.
The most important insight is your own PoV may be flawed. The way a scientist provides value via software is different than how a firmware developer provides value.
Was trying to remember a counter example on good hires and wasted money.
Alex St. John Microsoft Windows 95 era, created directX annnnd also built an alien spaceship.
I dimly recalled it as a friend in the games division telling me about some someone getting 5 and a 1 review scores in close succession.
Facts i could find (yes i asked an llm)
5.0 review:
Moderately supported. St. John himself hosted a copy of his Jan 10, 1996 Microsoft performance review on his blog (the file listing still exists in archives). It reportedly shows a 5.0 rating, which in that era was the rare top-box mark.
Fired a year later:
Factual. In an open letter (published via GameSpot) he states he was escorted out of Microsoft on June 24, 1997, about 18 months after the 5.0 review.
Judgment Day II alien spaceship party:
Well documented as a plan. St. John’s own account (quoted in Neowin, Gizmodo, and others) describes an H.R. Giger–designed alien-ship interior in an Alameda air hangar, complete with X-Files cast involvement and a Gates “head reveal” gag.
Sunk cost before cancellation:
Supported. St. John says the shutdown came “a couple of weeks” before the 1996 event date, after ~$4.3M had already been spent/committed (≈$1.2M MS budget + ≈$1.1M sponsors + additional sunk costs). Independent summaries repeat this figure (“in excess of $4 million”).
So:
5.0 review — moderate evidence
Fired 1997 — factual
Alien spaceship build planned — factual
≈$4M sunk costs — supported by St. John’s own retrospective and secondary reporting
Although, to be fair this is far from vibecoding. Your setup, at a glance, says a lot about how you use the tools, and it's clear you care about the end result a lot.
You have a PRD file, your tasks are logged, each task defines both why's and how's, your first tasks are about env setup, quality of dev flow, exploration and so on. (as a nice tidbit, the model(s) seem to have caught on to this, and I see some "WHY:" as inline comments throughout the code, with references to the PRD. This feels nice)
It's a really cool example of "HOW" one should approach LLM-assisted coding, and shows that methods and means matter more than your knowledge in langx or langy. You seem to have used systems meant to help you in both speed of dev and ease of testing that what you got is what you need. Kudos!
I might start using your repo as a good example of good LLM-assisted dev flows.
That seems a little bit dangerous, why not do it in a language you know ? Plus, this is not launching rockets on the moon, it's a sentence splitter with a fancy state machine (probably very useful in your niche, not a critique) - the difficulty was for you to put the effort to build a complicated state machine, the rest was frankly... not very LLM-needing and now you can't maintain your own stuff without Nvidia burning uranium.
Did the LLM help at all in designing the core, the state machine itself ?
Nah it was a hobby project because I was laid off for a bit.
Rust's RegEx was perfect because it doesn't allow anything that isn't a DFA. Yes-ish, the LLM facilitated designing the state machine, because it was part of the dev-loop I was trying out.
The speed is primarily what enabled finding all of the edge cases I cared about. Given it can split 'all' of a local project gutenberg mirror in < 10 seconds on my local dev box I could do things I wouldn't otherwise attempt.
The whole thing is there in the ~100 "completed tasks" directory.
I made this to try out Claude code. I have the $17/mo thing, and i don’t know rust. (I do know plenty of other languages.) Rust felt like a scripting language when used this way. I used a task system to force getting to a git commit before auto-compact. The completed tasks are in the repo allowing one to see what starting context kicked off the changes. It worked much of the time. It’s 8k of rust, 12k of markdown and i think the markdown helps to correctly interact with a codebase using agents just as unit tests assist in refactoring. On my I9 with a local gutenberg mirror this e2e discovers 20k+ english novels, splits them into sentences, normalizes the sentences, keeps the origination text span and writes it out as tsv’s. It takes 7 seconds to complete that for 7 Gb of novels. Most importantly it splits the sentences the way i needed for the start of my pipeline.
Definitely interested if anyone find cases where it mis-splits english text from a novel.
I used it for 2 weeks with the cheap $17/mo sub. It is equal parts amazing, and frustrating.
I ended up with 8k lines of rust and 12k lines of markdown. I think those markdown designs and explicit tasks were required the same way unit tests with a test harness are required to make the human-tool interaction work.
However, I’m not sure if the ‘magic’ is VC-subsidy or something else.
It did make rust (a language i do not know) feel like a scripting language. … the github repo is ‘knowseams’.
I loved discovering that rust has O(n) guardrails on regex! The so-called features that break that constraint are anti-features.
Over the last two weeks I wrote a dialog aware english sentence splitter using Claude code to write rust. The compile error when it stuck lookarounds in one of the regex’s was super useful to me.
And it is a developer feature hidden from end users. e.g. - In your ollama example, does the developer ask end users to install ollama? Does the dev redistribute ollama and keep it updated?
The ONNX format is pretty much a boring de-facto standard for ML model exchange. It is under the linux foundation.
The ONNX Runtime is a microsoft thing, but it is an MIT licensed runtime for cross language use and cross OS/HW platform deployment of ML models in the ONNX format.
That bit needs to support everything because Microsoft itself ships software on everything.(Mac/linux/iOS/Android/Windows.
ORT — https://onnxruntime.ai
Here is the Windows ML part of this —https://learn.microsoft.com/en-us/windows/ai/new-windows-ml/...
The primary value claims for Windows ML (for a developer using it)— This eliminates the need to: Bundle execution providers for specific hardware vendors
Create separate app builds for different execution providers
Handle execution provider updates manually.
Since ‘EP’ is ultra-super-techno-jargon:
Here is what GPT-5 provides:
Intensional (what an EP is)
In ONNX Runtime, an Execution Provider (EP) is a pluggable backend that advertises which ops/kernels it can run and supplies the optimized implementations, memory allocators, and (optionally) graph rewrites for a specific target (CPU, CUDA/TensorRT, Core ML, OpenVINO, etc.). ONNX Runtime then partitions your model graph and assigns each partition to the highest-priority EP that claims it; anything unsupported falls back (by default) to the CPU EP.
Extensional (how you use them) • You pick/priority-order EPs per session; ORT maps graph pieces accordingly and falls back as needed. • Each EP has its own options (e.g., TensorRT workspace size, OpenVINO device string, QNN context cache). • Common EPs: CPU, CUDA, TensorRT (NVIDIA), DirectML (Windows), Core ML (Apple), NNAPI (Android), OpenVINO (Intel), ROCm (AMD), QNN (Qualcomm).