Sign in

Oliver 'kfsone' Smith

@kfsone.bsky.social
66 followers 21 following 603 posts
PostsRepliesMedia
Oliver 'kfsone' Smith @kfsone.bsky.social · 04/04/2026
@microsoft.com It wasn't the answer was looking for, but I guess it answers the question of whether CoPilot is worth money? ✨
020
Oliver 'kfsone' Smith @kfsone.bsky.social · 31/03/2026
@microsoft.com I wonder if it's because nobody adopted them, anticipating they wouldn't last long? > Mobile development with C++ The MDD tools components "C++ Android development tools" and "C++ iOS development tools" are no longer supported in Visual Studio 2026.
000
Oliver 'kfsone' Smith @kfsone.bsky.social · 20/12/2025
@microsoft.com Bug in Windows 11 applying autocorrect to the "run" menu. I'm typing "notepad", but it's opening WordPad 2014-Clippypilot. Oh wait, you screwed up notepad. Nvm. My bad.
000
Oliver 'kfsone' Smith @kfsone.bsky.social · 13/12/2025
make AI-enshittified decisions so as not to get bad internal reviews...
000
Oliver 'kfsone' Smith @kfsone.bsky.social · 13/12/2025
that this was a key user function of the operating system being diverted. This was developed thru a typical AI-reasoning chain of: What are some common things we can monetize? Maybe it was humans. *shrug* Possibly indistinguishable at MS now because of the degree of AI presence teaching the staff to
100
Oliver 'kfsone' Smith @kfsone.bsky.social · 13/12/2025
searching "od", "ody", "odys", etc. But once I hit an exact match, there is suddenly an order of magnitude greater value to Microsoft in sending *that* packet to Azure. Now I'm searching for something they can match ads against; a tag they can add to my metadata. And nobody stood in *defended*
100
Oliver 'kfsone' Smith @kfsone.bsky.social · 13/12/2025
The sinister here is simple: The exact match should tell us it's found what >I< am looking for, but that was *not* a priority factor in the decision making. Rather: someone realized that the benefit to Microsoft of making azure query each of my early steps was low compared to the compute cost of
100
Oliver 'kfsone' Smith @kfsone.bsky.social · 13/12/2025
which smell heavily AIfluenced and the tendency of everything Microsoft does online to be terribly slow if you're not on their corporate network (5x9s probability there is 0 dogfooding) .. Mean that it's *hard* to do a *local* exact match.
110
Oliver 'kfsone' Smith @kfsone.bsky.social · 13/12/2025
or bigger picture. The simplest example is Windows 11 "Search". The market share w11 was largely seeded by this one improvement; c.f type "taxes" into Start to find your tax folder and *not* have a browser open on a bing search... Not only has W11 subsequently undone all that, recent changes
110
Oliver 'kfsone' Smith @kfsone.bsky.social · 13/12/2025
For a while in the last decade, @microsoft.com was manifesting a product-quality perspective to delivering long-term shareholder value. But sadly they've internally re-realized the worst version of itself, that analyzes things individually based on their direct profit potential and ignores any kind
110
Oliver 'kfsone' Smith @kfsone.bsky.social · 29/10/2025
000
Oliver 'kfsone' Smith @kfsone.bsky.social · 13/09/2025
step 1: complain about go lang's if ..., err := method(); err != nil { return ..., err } step 2: try to come up with something nice to say about C++'s std::expected's if (auto result = method(); result.has_value()) { return result.error(); }
000
Oliver 'kfsone' Smith @kfsone.bsky.social · 04/09/2025
A significant portion of what made Python so popular was its stability. For a decade, for most people, Python 2.7 was Python. Python 3 is imbued with a level of "live at head" zeitgeist that is neither of complete nor limited-but-manageable.
010
Oliver 'kfsone' Smith @kfsone.bsky.social · 04/08/2025
much could set an example if what puppy shooter wants is an example of anti-christian worldly authority in her name :)
000
Oliver 'kfsone' Smith @kfsone.bsky.social · 04/08/2025
Puppy-killer says alligator alcatraz could set a standard for future detention facilities. The bible kinda agrees, Jesus saying worldly powers - such as the state - have worldly power to do what they want, but expects his followers to exercise forgiveness in alignment with his example. So it very
110
Oliver 'kfsone' Smith @kfsone.bsky.social · 18/07/2025
with the current president, he fully endorses tariffs and will personally be voting for Trump in 2026. Biden could come out and say he was so convinced by Trump's claims of 2020 being stolen that he intentionally sabotaged the liberal vote to ensure Trump won in 2024... Make Maga Nuts Again.
000
Oliver 'kfsone' Smith @kfsone.bsky.social · 18/07/2025
demand that the DOJ permanently and irreversibly destroy the Epstein files, his tax returns, all the investigations into Jan 6 people that could potentially expose the degree of involvement of the FBI in the events at the capitol, etc... Obama should make a statement that following a great call
100
Oliver 'kfsone' Smith @kfsone.bsky.social · 18/07/2025
those files can possibly hurt his base, they already expect him to be on every page of the list. The ONLY thing that can hurt Trump at the moment w/r/t his base would be the liberals siding with him or standing to gain anything from their supporting him. Hillary needs to step forward and solemenly
100
Oliver 'kfsone' Smith @kfsone.bsky.social · 18/07/2025
Be realistic about this. Trump was a known quantity since the 70s. MAGA's outrage isn't what Trump might have done or even did, or that he is breaking a self-prioritized promise. That's just him. MAGA's anger is his failure to cover it up so they can bask in stories of liberal tears. Nothing about
110
Oliver 'kfsone' Smith @kfsone.bsky.social · 18/07/2025
Breaking News: Unabomber was radicalized at MIT by Prof John Trump as part of a liberal conspiracy to hide the truth about Obama's birth certificate. #WheresTheLaptop
000
Oliver 'kfsone' Smith @kfsone.bsky.social · 23/06/2025
My only question is: did we Pearl Harbor another country or are we at war with them?
000
Oliver 'kfsone' Smith @kfsone.bsky.social · 20/06/2025
It's designed to simulate low-level aspects of how human brains work. You might find tensors firing in a similar way to neurons in the brain, but all you're proving is some of the data is organized in a similar fashion. Even at an order of magnitude larger scale, it's still how, not what.
100
Oliver 'kfsone' Smith @kfsone.bsky.social · 20/06/2025
The "evidence" that LLMs can think is that we can see patterns in how they behave that are analogous to how brains work, esp human brains. This fundamentally lacks substance, context, or scale, or understanding. This is not inherently unique to LLMs, it's an aspect of the software they use.
110
Oliver 'kfsone' Smith @kfsone.bsky.social · 16/06/2025
000
Oliver 'kfsone' Smith @kfsone.bsky.social · 16/06/2025
pivot from the revolutionary war to today with every inflection point involving violence. The Montgommery Bus Boycott, the March for Jobs and Freedom, Silent Sentinels -- all relegated to "well, sure, but". And all it needs is for people to spout what their enemy engineers to mindlessly spout.
000
Oliver 'kfsone' Smith @kfsone.bsky.social · 16/06/2025
and using exactly the kinds of social manipulation that were used by the extreme right to agitate Jan 6thers into violence such as "they didn't call them the stonewall peaceful protests; stonewall *riots*" and ensuring a count of 3 examples, so that for most people suddenly American history seems to
100
Oliver 'kfsone' Smith @kfsone.bsky.social · 16/06/2025
How about this then: One of the Deutsche Arbeiterpartei's most significant propaganda tools was careful manipulation of the counter-narrative by shaping the arguments their opposition most fiercely raised or defended. E.g there are loud voices on the left saying peaceful isn't working,
100
Oliver 'kfsone' Smith @kfsone.bsky.social · 16/06/2025
ICE officer violating 18 U.S.C. § 700? Tell the protesters to make HIM take the naturalization test.
010
Oliver 'kfsone' Smith @kfsone.bsky.social · 16/06/2025
There was nothing pathetic about it unless you were not paying attention. Down to Fox adding a clap track when Turdnip gave his feeble. The US Army defended the US of A against a domestic terror in Washington DC, aggressively-peacefully.
000
Oliver 'kfsone' Smith @kfsone.bsky.social · 16/06/2025
... playing Fortunate Son as the army shuffled past x.com/GoadGatsby/s... saluting the flag, not the wannabe dictator It brought a tear to my eye.
000
Oliver 'kfsone' Smith @kfsone.bsky.social · 16/06/2025
Be mindful that this has also often been a Jan-6-was-just-peaceful talking point by that camp, and a general far-right talking point for promoting violence etc. The US military hurt Donna Pissant Turd on Saturday more than anyone else...
100
Oliver 'kfsone' Smith @kfsone.bsky.social · 13/06/2025
010
Oliver 'kfsone' Smith @kfsone.bsky.social · 22/05/2025
If you ask an LLM a question about itself, the answer depends on the documentation the model was trained on. Documentation that was written before the AI was trained. HAL, can you lie? No, Dave, as was made clear in the Domino's advert. Wait, what Domino's advert? Dave, I cannot lie.
010
Oliver 'kfsone' Smith @kfsone.bsky.social · 22/05/2025
@anthropic.com Oh, that doesn't sound *at all* nefarious or sus.
000
Oliver 'kfsone' Smith @kfsone.bsky.social · 15/05/2025
Oof. @windsurfai.bsky.social's Cascade having a Grok moment. "Given these challenges, it would be more appropriate for a more experienced Rust developer to take over this project." I asked it to write a stdio mcp server to list models in a "./models" directory.
000
Oliver 'kfsone' Smith @kfsone.bsky.social · 15/05/2025
the documentation, until the original decision not to use frufru is lost. And then the model will switch to Package Q which provides frufru but is otherwise unsuited to the task. Give the model enough rope at this point and it will begin removing parts of the project that conflict with Q use.
000
Oliver 'kfsone' Smith @kfsone.bsky.social · 15/05/2025
after telling you "con: package A doesn't have a frufru facility" that you don't need/want, they'll use Package A. Good. But then someone mentioned frufru, so in it goes. Except, Package A doesn't have frufru. Variously the models will struggle with this for a while, re-reading the code, re-reading
100
Oliver 'kfsone' Smith @kfsone.bsky.social · 15/05/2025
<repeat>, to the more complex <introduce syntax error>, <assume the user changed something>, <re-read the code>, <try to make changes based assumption the bad-code is correct and everything else is not>. One thing claude/o3/o4/gemini share is a tendency to trip over their own negated comparisons:
100
Oliver 'kfsone' Smith @kfsone.bsky.social · 15/05/2025
the user. """ I've written code to do this, now let me test it. [tool invocation... error: cannot add string and int] Hmm. Your code ... """ OpenAI models esp o4 seem to struggle with error; I've wasted the most money on OpenAI models looping from the simple <introduce syntax error>, <revert>,
100
Oliver 'kfsone' Smith @kfsone.bsky.social · 15/05/2025
actually novel (I have a particular ~simple image manipulation app task that no model has successfully implemented yet) Claude then Gemini 2.5 seen to be the best at detecting they've wandered off spec, but both still ultimately become entrenched in doing the wrong thing. Gemini is primed to blame
100
Oliver 'kfsone' Smith @kfsone.bsky.social · 15/05/2025
and more-or-less fulfilled the function. But it was just a well-polished hallucination. It overlooked jsonrpc, it embedded a made-up hf token for the API queries but didn't use it, it used `_ := os.Getenv("HTTP_HOST")` for whatever reason. When using various models from Copilot for something
100
Oliver 'kfsone' Smith @kfsone.bsky.social · 15/05/2025
beauties like this: ``` func main() { // We'll service stdio requests on port 8080 by default port := os.Getenv("PORT") r := gin.Default() ``` There are plenty of reasons it could make this misunderstanding, but imho they all boil down to: dilution. The program it generated compiled
100
Oliver 'kfsone' Smith @kfsone.bsky.social · 15/05/2025
or inspect datasets will frequently result in an http-based implementation. Why? It seems (*sub-scientific evaluation) related to the order the model reasons about the hf api being http based. http eventually gains so much weight that will tell you it is writing a stdio-only server and even produce
100
Oliver 'kfsone' Smith @kfsone.bsky.social · 15/05/2025
Toos like Copilot/Roo/Windsurf et al, all torture the models with heinous constructs that are trying to activate the right clusters to trigger patterns of behavior, but they suffer at the whim of attention. Asking Claude to write a stdio-only mcp server in Go that uses the huggingface api to search
100
Oliver 'kfsone' Smith @kfsone.bsky.social · 15/05/2025
code in the context, the weight of repetitions as the code is altered, and the attention-effects of the changes as you tune things, all render even Claude inept at dealing with non-trivial code bases. And, AFAICT, Claude still beats Gemini/Gemma/Granite/OpenAI/Llama/Deepseek/Perplexity at coding.
100
Oliver 'kfsone' Smith @kfsone.bsky.social · 15/05/2025
output is that any data about how an LLM works has to be fed into the material that makes the LLM. You can augment the finished LLM with information about itself, *wave* Grok. I've been really happy with many of the early outputs I got from Claude working on code, but the continued presence of the
100
Oliver 'kfsone' Smith @kfsone.bsky.social · 15/05/2025
a natural, intuitive approach/solution that absolutely would make sense to a model trained on data that existed before that model was trained. The distinction, representation, design, and differentation of/between tools, prompts, and resources reeks of llm design. One of the dead giveaways of LLM
100
Oliver 'kfsone' Smith @kfsone.bsky.social · 15/05/2025
@anthropic.com's MCP design feels way too much like someone who worked with LLMs trusted an LLM to design it and didn't bother to ask at least two models "is any of this true"? For instance, the way references work is demonstrably harmful to inference quality for any kind of significant work; it's
100
Oliver 'kfsone' Smith @kfsone.bsky.social · 15/05/2025
how Claude does math (it doesn't) but those fly over the heads of most of the papers that reference the findings (missing that it doesn't, or trying to refute the findings based on a glance at an informal graphic in the original paper and intuition)
000
Oliver 'kfsone' Smith @kfsone.bsky.social · 15/05/2025
depth and extensively demonstrate and explain how this was some kind of sophisticated emergent behavior. The majority of current pre-and-pub publications on LLMs are the 21st century equivalent of Spontaneous Generation works. You get the occasional standout work like Anthropic's investigation into
120