unkz's recent activity
-
Comment on Outsourcing plus local AI will soon become more economical vs frontier labs in ~tech
-
Comment on I think Anthropic and OpenAI have found product-market fit in ~tech
unkz Link ParentPeople used to be able to make an argument that compilers were less efficient than hand tuned machine code, but nowadays, I’m pretty confident that all in all but the most unusual circumstances an...People used to be able to make an argument that compilers were less efficient than hand tuned machine code, but nowadays, I’m pretty confident that all in all but the most unusual circumstances an optimizing compiler will far outperform a human.
What this means I think is that our best efforts from our brightest minds in the machine language space are now spent on optimizing compilers instead of writing bespoke machine code to solve individual problems.
This is what I see in the future: we will still have high level programming experts and mathematicians who know C or Rust or high dimensional mathematics in a deep way, but their efforts will be largely focused on improving coding agents who then in turn interface with “regular” developers, like compilers or interpreted languages do today.
-
Comment on I think Anthropic and OpenAI have found product-market fit in ~tech
unkz Link ParentThis is an interesting problem for sure, and I’ve spent a lot of time thinking about it. On the other hand, we used to say that about assembly language programmers.This is an interesting problem for sure, and I’ve spent a lot of time thinking about it. On the other hand, we used to say that about assembly language programmers.
-
Comment on What are your personal crackpot conspiracy theories about the world right now? in ~talk
-
Comment on What are your personal crackpot conspiracy theories about the world right now? in ~talk
unkz Link ParentI’ve gotten the last 6 digits of my phone number several times, which really just seems impossible.I’ve gotten the last 6 digits of my phone number several times, which really just seems impossible.
-
Comment on Weekly US politics news and updates thread - week of May 25 in ~society
unkz Link ParentAm I OOL here? Who was arrested besides Maduro?the second bogus arrest of a sovereign country's leader
Am I OOL here? Who was arrested besides Maduro?
-
Comment on Weekly US politics news and updates thread - week of May 25 in ~society
unkz Link ParentWhat does this mean?(but like was mostly pissed at being told what to do and he has the rights to the name?)
What does this mean?
-
Comment on Weekly US politics news and updates thread - week of May 25 in ~society
unkz LinkThis is one of the more offensive things I have seen so far. https://aliens.gov No, it’s not about Area 51. It’s just racism. It is interesting that they straight up admit that illegal immigrants...This is one of the more offensive things I have seen so far.
No, it’s not about Area 51. It’s just racism.
For 60 years, the U.S. government has kept a closely guarded secret.
Aliens have been walking among us, living in our neighborhoods, and interacting with us in our daily lives.
They've shopped in the same stores, attended the same classes as our children, and lived seemingly normal human existences.
With one exception — they do not belong here.
It is interesting that they straight up admit that illegal immigrants are just like every other American. No more pretense about only deporting violent offenders.
-
Comment on Who else is as excited as I am for the Backrooms movie tomorrow? in ~movies
unkz LinkI know my kids are pretty hyped, so I imagine I’ll be watching it one way or another.I know my kids are pretty hyped, so I imagine I’ll be watching it one way or another.
-
Comment on Actually useful MCPs in ~comp
unkz (edited )Link ParentBasically I have it hooked up to openclaw and obsidian, through a skill that calls Hevy MCP endpoints and the standard obsidian skill. I have five different workouts in Hevy labelled...Basically I have it hooked up to openclaw and obsidian, through a skill that calls Hevy MCP endpoints and the standard obsidian skill. I have five different workouts in Hevy labelled Monday-Friday. I have an obsidian document that contains my overall goals, lists of acceptable exercises. I’m integrated with telegram so every morning I get a message telling me what my workout is going to be, which is really just for fun, then I go to the gym, do the workout, sometimes leaving notes like “very easy, should increase weight”, “skipped today because rushed for time”, etc. It’s set up to poll until my workout is done and then it analyzes, proposes changes (increase reps to 12, increase weight to 60 and drop reps to 8, schedule missed exercise for Friday, etc).
Useful stuff it’s done for me lately include reorganizing my schedule around an injury and my physiotherapist’s prescribed rehab exercises, and planning out a phased return to the gym after being out for two weeks with a tonsil infection.
On occasion I’ll find that it’s doing something like scheduling too much or not enough and I’ll have to chat a bit about it but generally it’s well behaved. I maintain my obsidian vault in git and it has a copy of all my schedule changes just in case but I haven’t needed that backup yet.
-
Comment on Actually useful MCPs in ~comp
unkz LinkNot for development but I integrate with Hevy to let an LLM manage my gym workouts. It monitors my progressive overload progression, plans around my other sports, if I miss a workout it’ll...Not for development but I integrate with Hevy to let an LLM manage my gym workouts. It monitors my progressive overload progression, plans around my other sports, if I miss a workout it’ll redistribute my exercises, things like that.
-
Comment on If you let AI do your writing, I will come to your house and kill you in ~tech
unkz Link ParentIn principle yes, but only in a research setting. It would be cost prohibitive to do at scale.In principle yes, but only in a research setting. It would be cost prohibitive to do at scale.
-
Comment on If you let AI do your writing, I will come to your house and kill you in ~tech
unkz Link ParentI'm pretty aware of how language models work, having been building and training them from scratch since long before GPT even existed. I think this conflates the training objective (next token...I'm pretty aware of how language models work, having been building and training them from scratch since long before GPT even existed.
because LLMs are not designed to understand the works they're trained on. They are designed to generate text that looks like the works they're trained on. That's it.
I think this conflates the training objective (next token generation) with the learned capability: a high dimensional model of the latent structure that produced the tokens: language, concepts, facts, social patterns, relationships, and many elements of humanity in general.
-
Comment on If you let AI do your writing, I will come to your house and kill you in ~tech
unkz Link ParentFrankly, i think you are wrong. Moby Dick is also not human — it’s a bunch of paper and ink, yet it offers insight into the human condition. Why? It is also an artifact of human intellect. That is...AI cannot offer insight into the human condition, because AI is not human.
Frankly, i think you are wrong. Moby Dick is also not human — it’s a bunch of paper and ink, yet it offers insight into the human condition. Why? It is also an artifact of human intellect. That is also what an LLM is, an artifact of human intellect, but one that has digested not only Moby Dick, but also millions of other artifacts of human intellect. Saying there is nothing to learn about the human condition from an LLM is like saying there is nothing to learn from going to the library.
-
Comment on If you let AI do your writing, I will come to your house and kill you in ~tech
unkz Link ParentI guess because I consume millions of tokens per day and I’m very familiar with the failure modes of these models. Hallucinations absolutely happen in complex conversations with long context...I guess because I consume millions of tokens per day and I’m very familiar with the failure modes of these models. Hallucinations absolutely happen in complex conversations with long context windows. This query though? The chances of it not doing a tool call to an external data provider to confirm is nil.
-
Comment on If you let AI do your writing, I will come to your house and kill you in ~tech
unkz Link ParentIt’s also actually impossible to do on large scale models because they are not in fact deterministic. I don’t mean that in the sense that they generate probability vectors, or that they rely on...It’s also actually impossible to do on large scale models because they are not in fact deterministic. I don’t mean that in the sense that they generate probability vectors, or that they rely on external tool invocations because the indexing into those vectors and the search results could be stored.
The problem is down to hardware and concurrency. Where a job gets scheduled matters, and different chunks of model getting executed in parallel means execution order varies and small floating point errors start to add up. To get true determinism means running specialized models on the exact same hardware under the exact same conditions.
That said, no current frontier model is going to hallucinate the way the author claims on such a basic and straightforward question.
-
Comment on If you let AI do your writing, I will come to your house and kill you in ~tech
unkz Link ParentIf only I had an extra 24 hours in my day.If only I had an extra 24 hours in my day.
-
Comment on If you let AI do your writing, I will come to your house and kill you in ~tech
unkz Link Parent… yes. It will be great when deep insight into the human condition can be turned like a tap. I would like to read the complete The Mystery of Edwin Drood, The Pale King, or an actual good ending...Is it better if/when AI can convincingly emulate a particular, high quality human author?
… yes. It will be great when deep insight into the human condition can be turned like a tap. I would like to read the complete The Mystery of Edwin Drood, The Pale King, or an actual good ending to Game of Thrones.
These articles seem to come from one of two main places, often simultaneously: a conceit that there is something about humanity that is ineffable and cannot be replicated by a machine, and a deep seated fear that we aren’t nearly as ineffable as we think. Either way, I think pretty soon we’re all going to be effed.
-
Comment on If you let AI do your writing, I will come to your house and kill you in ~tech
unkz LinkI keep reading articles like this which seem to come from a parallel universe that is 12-18 months in the past. I asked ChatGPT in its shittiest quality level (instant) and got exactly what I...This is why you can still ask an AI to tell you about the scene in VS Naipaul’s Dashed Against the Rocks in which a donkey is thrown from a hot air balloon, and it’ll tell you that ‘what matters in Naipaul’s handling is not the event’s cruelty in isolation, but its emotional flatness and the sense of mismanaged modernity,’ despite the fact that there is no scene in Dashed Against the Rocks where a donkey is thrown from a hot air balloon, and also no novel by VS Naipaul called Dashed Against the Rocks.
I keep reading articles like this which seem to come from a parallel universe that is 12-18 months in the past.
I asked ChatGPT in its shittiest quality level (instant) and got exactly what I expected:
I can’t find any V. S. Naipaul work titled Dashed Against the Rocks, nor any well-known Naipaul scene involving a donkey being thrown from a hot-air balloon. I checked references to Naipaul’s fiction and nonfiction, and nothing matching that episode appears in the standard bibliographies or criticism.
Did they actually ask an LLM this question? Or did they imagine what an LLM would say based on their hazy memories of what an LLM once told them in 2025, or cobble together this response from recollections of other articles by other people who are angry about AI? Dare I ask — did they hallucinate this response?
-
Comment on Project Glasswing: An initial update in ~tech
unkz LinkI’m curious to see if flyingpenguin will have a response to this, if they will move their goalposts now that details from multiple projects are released, or if they will quietly ignore it.I’m curious to see if flyingpenguin will have a response to this, if they will move their goalposts now that details from multiple projects are released, or if they will quietly ignore it.
https://developers.openai.com/cookbook/examples/reproducible_outputs_with_the_seed_parameter
With the caveat that this is only mostly reproducible due to fundamental issues with parallelism. It’s a best effort but on relatively short responses you won’t find too much drift.