Preface: I do think the nature of the work is changing, now that LLMs are capable code generators. And I'll admit that I started working with LLMs because of FOMO, but also to keep my skills sharp in an evolving market. I'm not 100% certain this will pan out, but I do think that working with LLMs is a skill unto itself, and I feel that I'm much better at it now than I was when I started (in earnest, roughly) a year ago.
Q: What's my setup like?
A: I'm using both Claude Code and opencode w/ DeepSeek flash. Recently I've started providing the task context to both simultaneously, and then have one act as a real-time reviewer, and one as implementer. I don't consistently use one or the other for implementer, and I'm still learning their capabilities and "who's good at what." But this allows me to get one model to provide an instantaneous code review to the other, as soon as the one model is done writing the code. I set my permissions for both to ask every time on every action: Read, Write, Bash, etc. I grew up playing turn-based RPGs, so I prefer turn-based battle to real-time.
Q: What kind of work I offload, vs do myself.
A: Almost all of the code is written by an LLM, but, since I'm reading and rejecting a lot of code, and basically micro-managing them, I feel that the code that's generated is roughly what I would have written myself. Work that I don't offload is any kind of email or communication, or non-code writing, with the exception of README files, and the prose session summaries I have them write.
Q: How do I make it enjoyable?
A: I make it enjoyable by finding new ways to work with the LLMs, and specifically, getting them to produce better quality code. The real-time code review process I outlined above is one method, but I've also found that the LLM will write better code when given a "guiding principles" document.
Once they started passing the blame, almost immediately, I would expect the first thing to be a new email from Gates saying “Bob, you’re doing this, Frank, you’re doing that, no excuses, get it done.”
Feature-flag evaluation: DISABLE_TELEMETRY, DO_NOT_TRACK, CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC, and DISABLE_GROWTHBOOK each
disable the feature-flag evaluation that Remote Control availability depends on. Unset the variable
wherever it’s set, in your shell environment or in the env block of a settings.json file, to use Remote Control.
I just discovered the same thing, that my local sessions were sent to claude.ai/code and stored there. I added:
I'm only running a few agents in parallel so far, but I don't let them write or read anything w/out my approval. This slows things down quite a bit, but it means that I have exactly the opposite of your problem of stopping them.
I've only started really thinking about developing my own harness, but one of the reasons would be to implement some mechanism where I could give the agent a "turn budget", say, 5 turns, when I can clearly see what the agent is going to be doing for the next few turns, without my needing to allow each turn to go forward.
But it sounds like the same thing would be useful for you as well, just in the other direction.
Could you say more about what "certain points" you're trying to get them to stop at?
SQLite doesn't have system-versioned / temporal tables, but a quick search turns up a fairly straightforward approach. Instead of overwriting existing rows, always write new rows with timestamps. https://www.ohnekontur.de/2024/02/19/unlocking-time-harnessi...
I'm thinking now of the hoops you have to jump through to edit a package.json file to update your dependencies, and thinking yeah, what if you could do: "UPDATE dependencies SET version='1.2' where name='madlib';"
It’s interesting to think of the LLM as a kind of “mirror”.
It reads to me like this person is engaging with the tool in a fairly hostile manner, and the LLM is like, “so you want to play it like that, huh? Well, let’s play it like that then!”
The language that you use to prompt the LLM is the language you’re going to get back.
Not necessarily. Opus used to become overly apologetic and started to second guess everything it did after getting insulted or "talked down", as recently as the beginning of the year. It was a surprisingly effective way to get it to behave for a while lol. I haven't seen that behavior since at least 4.8 though.
My personal opinion: all output from LLMs is hallucinations. The idea that it’s “right” or “wrong”, when stating factual information is really in the eye of the beholder/user. What year did astronauts land on the moon? — is a question that may or may not have a factual answer depending on your own beliefs. Philosophically, you would need to also define what a fact is, or what “not hallucinating” is, to define what a hallucination is. My understanding is that this property of the LLM architecture is innate, and until we have “world model” LLMs, or some other model that reasons from first principles, instead of the current “guess the next word” model, this isn’t going away. Just don’t rely on the for facts.
Is this distinct from the hardware in a land mine making the decision to explode shrapnel through a human being without any human in the chain except the dead guy.
Worth noting that both cases indirectly involve the humans that designed devices and the humans that made the placement and trigger condition decisions.
Further:
> AI drones are being used to autonomously target and kill targets by the Ukraine using technology they have been given.
are remote operated, they allow defenders cover while themselves being out and exposed.
However were they altered to autonomously fire, that would be on the basis of pattern matching in the visible and infra red spectrum - shoot at all hot blobs.
That's more of a trigger threshold setting issue than an LLM hallucination issue, and the danger is on par with any weapon system on auto fire, you really shouldn't approach such things until they are put in a safe off state or have exhausted ammunition.
I guess what I'm really asking is whether or not it's moral and ethical to use AI that targets autonomously? My understanding is that it's being used that way, but I can't back that up and I'll take your word for it if that's not the case. But it is certainly a plausible scenario. Everyone invested in AI seems determined to put it in everything because they need consumers to want to pay for it, and right now they haven't figured out a way to earn back from consumers the over a trillion dollars that has been put into the AI bubble. I don't have much faith that the people with billions invested don't want return on their investment, and I don't think they care about the consequences necessary to get it back.
> moral and ethical to use AI that targets autonomously?
First point, vision systems have been used in industry to look for misaligned labels, incorrectly folded papers (in high speed paper presses), wrong items on high speed conveyor belts etc. for thirty odd years now - they have issues that a very distinct from LLM 'AI' hallucinations.
That's nomenclature out the way.
Landmines are indiscriminate, they trigger on any weight or pressure over a threshold.
A vision based Felixer, by contrast, only triggers on cats (well, almost always only) and leaves bilbies and bettongs to walk on by.
That's an improvement over landmines.
The crux of your issue here might be the morality and ethics of establishing human exclusion zones within which all humans are highly likely to die.
These historically are created with rapid patterned artillery fire, butterfly mines, Napalm, indiscriminate criss crossing machine gun fire, etc.
Now there exists an option to use drones to kill all humans and leave the horses and cows alive.
Is it your concern that a bad vision threshold might kill a horse rather than a person? (Likely not)
Would you prefer an area to be napalm'd and agent orange'd back to dust?
Also, why do people keep talking about landmines? I'm talking about software, you are talking about dumb mechanical things from WWII. This isn't a philosophical question about methods of warfare from the past, but the software that controls the dumb hardware.
I, a person, singular, mention landmines as they are a hardware device that are designed to trigger action on a threshold.
Felixers are computer vision based devices that are designed to trigger action on a threshold.
These are actual real world objects that Bishop Berkeley can kick, not vague philosophical questions but actual engineered devices that kill and are in use today.
I don’t trust their outputs yet enough to not review everything myself. But I’m happy to be. a QA bottleneck in that way. What I want now is orchestration to get them running in parallel. I’ve looked into Yegge’s beads, but I want something similar that is far simpler, fewer sloc, and database-driven rather than json-file-driven. Probably need to build it myself with a graph database…
Preface: I do think the nature of the work is changing, now that LLMs are capable code generators. And I'll admit that I started working with LLMs because of FOMO, but also to keep my skills sharp in an evolving market. I'm not 100% certain this will pan out, but I do think that working with LLMs is a skill unto itself, and I feel that I'm much better at it now than I was when I started (in earnest, roughly) a year ago.
Q: What's my setup like? A: I'm using both Claude Code and opencode w/ DeepSeek flash. Recently I've started providing the task context to both simultaneously, and then have one act as a real-time reviewer, and one as implementer. I don't consistently use one or the other for implementer, and I'm still learning their capabilities and "who's good at what." But this allows me to get one model to provide an instantaneous code review to the other, as soon as the one model is done writing the code. I set my permissions for both to ask every time on every action: Read, Write, Bash, etc. I grew up playing turn-based RPGs, so I prefer turn-based battle to real-time.
Q: What kind of work I offload, vs do myself. A: Almost all of the code is written by an LLM, but, since I'm reading and rejecting a lot of code, and basically micro-managing them, I feel that the code that's generated is roughly what I would have written myself. Work that I don't offload is any kind of email or communication, or non-code writing, with the exception of README files, and the prose session summaries I have them write.
Q: How do I make it enjoyable? A: I make it enjoyable by finding new ways to work with the LLMs, and specifically, getting them to produce better quality code. The real-time code review process I outlined above is one method, but I've also found that the LLM will write better code when given a "guiding principles" document.
HTH
reply