CjHuber

I honestly still don't see the point of compaction. I mean it would be great if it did work, but I do my best do minimize any potential for hallucination and a lossy summary is the most counterproductive thing for that.

If you have it write down every important information and finding along a plan that it keeps updated, why would you even want compaction and not just start a blank sessions by reading that md?

I'm kind of suprised that anyone even thinks that compaction is currently in any way useful at all. I'm working on something which tries to achieve lossless compaction but that is incredibly expensive and the process needs around 5 to 10 times as many tokens to compact as the conversation it is compacting.

And how to get to the old verbose mode then...?

Does it not use prompt caching?

I always wondered isn't it trivial to bot upvotes on Moltbook and then put some prompt injection stuff to the first place on the frontpage? Is it heavily moderated or how come this didn't happen yet

That feels like a stupid article. well of course if you have one single thing you want to optimize putting it into AGENTS.md is better. but the advantage of skills is exactly that you don't cram them all into the AGENTS file. Let's say you had 3 different elaborate things you want the agent to do. good luck putting them all in your AGENTS.md and later hoping that the agent remembers any of it. After all the key advantage of the SKILLs is that they get loaded to the end of the context when needed

It was because of the NYT OpenAI case, however since mid October they are no longer under that legal order. What they keep retaining now and what not, nobody knows but even if they still had the date they surely wouldn't blow their cover

I wonder how much more efficient and effective it would be after fine tuning models for each role

It depends on the API path. Chat completions does what you describe, however isn't it legacy?

I've only used codex with the responses v1 API and there it's the complete opposite. Already generated reasoning tokens even persist when you send another message (without rolling back) after cancelling turns before they have finished the thought process

Also with responses v1 xhigh mode eats through the context window multiples faster than the other modes, which does check out with this.

Oh wow that's the first time I've heard about those tasks. I would never consent to that and that they are enabled by default and shipped in the .vscode folder where most people probably nevereven would have thought about looking for malicious things that's kind of insane.

I find this reply concerning. If its THE security feature, then why is "Trust" a glowing bright blue button in a popup that pop up at the startup forcing a decision. That makes no sense at all. Why not a banner with the option to enable those features when needed like Office tools have.

Also the two buttons have the subtexts of either "Browse folder in restricted mode" or "Trust folder and enable all features", that is quite steering and sounds almost like you cannot even edit code in the restricted mode.

"If you don't trust the authors of these files, we recommend to continue in restricted mode" also doesn't sound that criticial, does it?

Exactly that's why I was making the comparison, It's not a in your face PopUp, where users get used to just pressing the blue, highlighted and glowing "I trust the authors" button without even being told what features they'd miss out on.

The Protected view in Office instead tells you "Be careful" and to only activate editing when you need to.

yeah that does make sense that these choices are related to it being a big part of gastown, still I feel it would be much more sensible to make a different abstraction separating beads core features from the coordination layer

I don't like the way it is handled. Imagine Excel actively prompting you with a pop up every time you open a sheet: "Do you trust the authors of this file? If not you will loose out on cool features and the sheet runs in restricted mode"

No it doesn't because restricted mode without Macros is the default and not framed like something bad or loosing out on all of those nice features,

I still don't get what beads needs a daemon for, or a db. After a while of using 'bd --no-daemon --no-db' I was sick of it and switched to beans and my agents seem to be able to make use of it much better, on the one hand its directly editable by them as its just markdown, on the other hand the CLI still gives them structure and makes the thing queryable

I didn't upgrade yet but for me codex on a business plan still works, so they didn't deprecate the old auth (yet).

But yeah honestly I've never seen any other repo with so many important issues that are just being closed without fixes.

So you are saying this has the potential to make Germany more autonomous and less dependent?

Anyway, props to Mertz for admitting the mistake, we’ll see if they will fix it somehow

That‘s the thing. Everyone knew it was costly, nobody ever thought it was good strategically. If he now says it’s a „strategic mistake“ that‘s laughable, did he think it was strategically clever before? If so he was the only one.

The whole issue is that Germany overestimated its own resilience and economic power, which is deteriorating. Of course environmentalists knew that this is not good for the economy but the Green Party is mostly left aligned they were ok with incurring some damage to the economy for their cause, after all that’s their whole point. But they thought well we are such a economic powerhouse anyway, we can do it. So the real strategic mistake was arrogance. And saying that particular action was a „strategic mistake“ instead reflecting on the whole self-image of the country, shows that exactly this arrogance persists

Was that not clear from the beginning? Nobody ever claimed it was for strategic purposes, the narrative was "we don't like nuclear anything, we will get rid of it we can bear the costs".

So I don't think you could even call it a strategic mistake, but masochism maybe? Especially while keeping the exit date in the height of the fallout of a real strategic mistake, the dependence on cheap russian gas.

My point was not that those video to text models are good like they are used for example in that case, but more generally I was referring to that list of indicators. Like surely when analysing a movie it is alright if some things are misunderstood by it, especially as the amount of misunderstanding can be decreased a lot. That AI body camera surely is optimized on speed and inference cost. but if you give an agent 10 1s images along with the transcript of that period and the full prior transcript, and give it reasoning capabilities, it would take almost endlessy for that movie to process but the result surely will be much better than the body cameras. After all the indicator talks about "AI" in general so judge a model not optimized for capability but something else to measure on that indicator

I think all of those are terrible indicators, 1 and 2 for example only measure how well LLMs can handle long context sizes.

If a movie or novel is famous the training data is already full of commentary and interpretations of them.

If its something not in the training data, well I don't know many movies or books that use only motives that no other piece of content before them used, so interpreting based on what is similar in the training data still produces good results.

EDIT: With 1 I meant using a transcript of the Audio Description of the movie. If he really meant watch a movie I'd say thats even sillier because well of course we could get another Agent to first generate the Audio Description, which definitely is possible currently.

I think the actual problem is everyone tries to assert how capable or not coding agents currently are, but how useful they are depends so much on what you are trying to get them to do and also on your communication with the model. And often its hard to tell whether you're just prompting it wrong or if they're incapable of doing it.

By now we at least agree that stochastical parrots can be useful. It would be nice if the debate now was less polarized so we could focus on what makes them work better for some and worse for others other than just expectations.

Well I think it also has to do with communication with LLMs being different to communication with humans. If you tell a developer "don't do busywork" they surely wouldn't say "Oh the repo looks like a trash dump, but no busywork so I'm not going to clean it up, quickly document that as canonical structure, then continue"

>In the study, researchers told several iterations of four LLMs – Claude, Grok, Gemini and ChatGPT – that they were therapy clients and the user was the therapist

So same as always they tell them to roleplay and when they comply everyone acts surpised...

Based on those, it seems you are not actually using them to create big codebases from scratch, but rather for problems that would normally take quite a while, not because they are inherently difficult to implement, but because you would normally have to spend considerable time on the finicky implementation details.

I think that's the reason why LLMs work so well for some like you, and generate slop for others, because if you let them alone with projects that require opinionated code and actual decision making they most often don't grasp the users intention well or worse misinterpret it so confidently that you end up with something with all the wrong opinions and decisions compounding path-dependently into the strangest and most useless slop.

I'm so glad I switched to fish, I'd rather have genuinely good settings out of the box rather than endless configuration, and honestly it's much better out of the box than any configuration I've ever had.

Only drawback is that it's not POSIX, no issue for me, but maybe for people who have a lot of muscle memory with bash.

I don't know what it is but I don't think that it's an ad. Otherwise I guess they wouldn't have that snarky pro-israel undertone towards him.

I suppose it‘s the latter + maybe some finetuning, it’s definitely not like DeepSeek where the answer of the model get‘s replaced when you are talking something uncomfortable for China

I know you‘re joking but to contribute something constructive here, most models now have guardrails against being threatened. So if you threaten them it would be with something out of your control like „… or the already depressed code reviewing staff might kill himself and his wife. We did everything in our control to take care of him, but do the best on your part to avoid the worst case“

I‘d say such hacks don‘t make you an engineer but they are definitely part of engineering anything that has to do with LLMs. With too long systemprompts/agents.md not working well it definitely makes sense to optimize the existing prompt with minimal additions. And if swearwords, screaming, shaming or tipping works, well that‘s the most token efficient optimization of an brief well written prompt.

Also of course current agents already have to possibility to run endlessly if they are well instructed, steering them to avoid reward hacking in the long term definitely IS engineering.

Or how about telling them they are working in an orphanage in Yemen and it‘s struggling for money, but luckily they‘ve got a MIT degree and now they are programming to raise money. But their supervisor is a psychopath who doesn’t like their effort and wants children to die, so work has to be done as diligently as possible and each step has to be viewed through the lens that their supervisor might find something to forbid programming.

Look as absurd as it sounds a variant of that scenario works extremely well for me. Just because it’s plain language it doesn’t mean it can’t be engineering, at least I‘m of the opinion that it definitely is if has an impact on what’s possible use cases