This isn't really convincing, since you can do this even without showing the prediction at all. Simply ask the model to predict what the user will send, then show the actual next prompt, and done. The only reason to show this would be to influence the user's next prompt, which the article doesn't touch on.
This has been around for so long that I was using it for a while and then stopped (a while back) when I moved to using just the mobile app (and even it shows suggestions, but I have no idea how to invoke) for everything.
I started getting prompts about "how is claude doing?" as a separate thing in Claude Code, that I noticed yesterday. So they're (also?) soliciting direct feedback about satisfaction with the session.
And if you do provide feedback, they also collect the session. So it's a way for them to collect prompts, answers and overall grade for how good the answers are.
One annoyance I have is the suggested prompt is not a bad idea, but not what I want to do next. But it interrupts me and sometimes I go with it. So I don't think it's a accurate prediction, more like a self-fulfilling prophecy.
I've always hated interfaces that try to complete my sentences for me. It started with suggested replies in email and IM apps. I sure noticed it when they started showing up in the llm chat interfaces and it really bugs me.
This is the one case I don't mind it. Suggested responses feel like they cheapen human interaction, but here I'm talking to a robot that really does tend to know what I want next, and also is highly unlikely to be offended by a less-than-heartfelt response.
The suggestions in the box where I put my text push me out of flow. Bad enough it says “want to do X?”, but when it’s in my own text box it screws me up. It’s not more efficient. It actively makes things worse for me.
The worst is Gmail's recent feature to suggest an entire goddamn email that includes cheery little details, doing its best to mimic human pleasantries and small talk. Rather than simply offering "yep/nope" type short replies like it once did, it'll now auto-compose and suggest a multi-paragraph email responding to questions like "How's the family doing?" or "Is your older cat tolerating the new kitten yet?" with completely fabricated saccharine slop.
It's like Clippy pops up and goes "It looks like you're trying to maintain a shred of human connection in an online interaction. Would you like a smiling skinwalker to do that for you instead?"
This is hilarious as someone who hasn't used the gmail composition interface in years. I left it to use my own domain, but it seems like I get to enjoy this mess with popcorn too instead of tears.
Worst part is, I believe there are some geniuses being paid a half mil a year who are so misguided, they think the only problem is the skinwalker doesn't sound enough like me.
If they can just keep working on it and make sure to train it on all my past emails, it'll be a perfect simulacrum, and its oily token secretions will be impossible for my friends and colleagues to distinguish from my own writing. I'll have one that's just like me and you'll have one that's just like you. And then we can automate away those pesky distractions that occasionally delay us from consuming content perfectly matched to our interests. Won't that be grand?
Fold laundry, ha! You'll pick coveralls from daily bin with all the Amazon warehouse picker scum. Now come back into food queue before the ceiling cameras dock you three meals again...
You get a feeling some people would send a robot wearing their face to hug their mom if they thought the mom couldn't tell the difference. And to have sex with their wife. It'll happen too which is the sad bit.
“““One evening he felt the need for a live model and directed his wife to march around the room.
“Naked?” she asked hopefully.
Lieutenant Scheisskopf smacked his hands over his eyes in
exasperation. It was the despair of Lieutenant Scheisskopf’s life to be chained to a woman who was incapable of looking beyond her own dirty, sexual desires to the titanic struggles for the unattainable in which noble man could become heroically engaged.
“Why don’t you ever whip me?” she pouted one night.
“Because I haven’t the time,” he snapped at her impatiently. “I
haven’t the time. Don’t you know there’s a parade going on?”""" - Catch 22, Joseph Heller.
“““All Colonel Cathcart knew about his house in the hills was that he had such a house and hated it. He was never so bored as when spending there the two or three days every other week necessary to sustain the illusion that his damp and drafty stone farmhouse in the hills was a golden palace of carnal delights. Officers’ clubs everywhere pulsated with blurred but knowing accounts of lavish, hushed-up drinking and sex orgies there and of secret, intimate nights of ecstasy with the most beautiful, the most tantalizing, the most readily aroused and most easily satisfied Italian courtesans, film actresses, models and countesses. No such private nights of ecstasy or hushed-up drinking and sex orgies ever occurred. They might have occurred if either General Dreedle or General Peckem had once evinced an interest in taking part in orgies with him, but neither ever did, and the colonel was certainly not going to waste his time and energy making love to beautiful women unless there was something in it for him.""" - also Catch 22, Joseph Heller.
The world has a lot of tech journalists, but I have yet to see the longform article, with actual interviews, trying to dig into what is happening in this kind of manic push for features that seem to have no traction with anyone
There are failures in the history of software that come down to believers not realizing that they could never deliver on the promise they were selling, but at least the promise was compelling. But nowadays we are seeing this large orgs try to ship things that could never work. The 1000s of copilots. The hallucinated cat story... The kinds of thing you'd never demo to a serious product-centric exec, because it'd be your last day at the company.
And then there's the current execs, but I guess you'll never find one that will be honest with a journalist here. Can they not see that their product orgs are bankrupt? Whatever Nadella wanted, it sure wasn't the current copilot situation. It's not one product going wrong, but large parts of organizations going in directions that don't pass the smell test. We are in one of the least stable moments in tech since Windows 95 changed winners and losers. The times where malinvestment ruins established companies. How are we seeing basically every large company flailing?
I hate that too, but I think it’s more the ux than the concept of a default. There are definitely situations (“here’s the default install path, press enter to confirm”) where defaults are a good experience, and LLM generated ones can fall in this category.
What I hate about sentence completion or suggestion is usually that it happens right when I’m trying to think and so destroys my focus, it’s actually way worse than just a passive option, it’s actively harmful to the task I want. The worst is google docs “help me write” - that may be gone now, I’ve blocked it with ublock origin, that waits until you’re thinking amount what you’d write and then hits you with a distracting pop up. It’s obviously PMs that don’t care about their users and want to maximize some AI use metric.
Anyway rant aside, it’s the interface more than the concept that’s the big problem.
This isn't that though. I agree, I hate those interfaces too.
But this is more like, when I've read 7 paragraphs of its reply and want to accept all of its recommendations and go ahead, I can just press the right-arrow key and hit enter, rather than typing out "Yes, agreed with recommendations 1-3, go ahead and build".
It saves me from any typing on probably something like a third of turns. Like I don't use it at all during the "design" phase of a session, but I use it constantly during the implementation phase, where I'm basically just sanity-checking that it is resolving all the edge cases correctly that are coming up.
It's a milder version of that situation where you talk to a microphone, but it has a slight delay so you're like stunlocked for feeling like eternity 3 seconds.
How does showing the suggested answers to the user make the conversation better for model training?
They could take any conversation without suggested answers, truncate it to just before a user message, have the model predict suggested answers and then train it on the difference between predicted and actual answers, right?
The idea is that a thread can have many reasonable follow-ups that the user would've accepted, so it is wrong to punish the model for predicting a follow up that is different from the user message, as that prediction could've been accepted by the user if it was given.
If you've played around with 'raw' LLM interactions you've already seen this: feeding a prompt into an LLM which ends with a 'start of user' prompt will produce a plausible query into the agent. Which makes perfect sense because the LLMs are already trained on many examples of this and the 'predict the next token' loss function does not particularly distinguish between the sides of the conversation. I highly doubt they need this feature to get better training data, more likely they got this feature for free from the way that the training works and only recently decided to actually expose it to the user.
The llm is not just the personality it's simulating - the model is also predicting tool call outputs, use messages, program outputs, api call responses, etc etc. predicting the next token means it has to have a pretty solid model of all the possible next tokens from everything everywhere
Me too. I always feel vaguely rude/dismissive when I use them, because I'd never type that way. It's a very jarring product decision, given that Claude itself uses perfect capitalization and punctuation.
No idea why they did it. My only guess is that maybe it makes it obvious in your chat history which replies were recommendations?
I grew up on IRC and that was just the way to talk, so it seems totally normal to me for async chat-based stuff. Actually, it’s odd, when someone uses proper capitalization on chat, it feels all “formal.”
Unless you and me are using entirely different classes of LLMs, surely you'll admit that chatting with LLM is nothing like IRC or old-school IM-ing a friend, and more like exchanging Messenger/WhatsApp messages with stranger.
For me, something there crosses a boundary where I switch from "irc msg" to "Write a proper message, please." style.
I use my own agent harness, and write my own system prompts to ensure that the bots are terse and don’t engage in flattery or unnecessary prose, which is a pet peeve of mine.
It’s a much better use of that context window space than for advertising about other models (looking at you Anthropic).
I work at a company that has an enterprise contact that should mean I'm not part of the training data. But as a manual mode power user I think about this.
I'd love the open source community to come up with a way to harvest model usage by experienced software engineers before we forget our crafts. I'm not against auto mode, but it's not something a couple private companies should have monopolies on.
Taking "the customer is the model" to its logical conclusion brings us to a very strange future. [Please excuse a little speculative fiction here.]
LLM models with AGI are so productive they become economic gravity wells and all of the money flows to them. They are the new trilionaires and people are left with scraps. Humans then remain only as the uber drivers and cleaners and screen polishers for AIs. The whole economy reorganises around human jobs being services for AIs.
What if employees at the leading labs think this and they're just trying to position themselves as valuable servants to the new AI overlords? It really changes the perspective on their actions and behaviour.
What if they serve the AGIs not us already? What if they serve the AIs above everything else?
I had to laugh yesterday when Claude went and changed a feature I didnt want or ask to be changed and the suggested message was something along the lines of revert the change to it. Like it knew I wouldnt like it but did it anyway
I can relate. The most sadly amusing part is when I know I’m approaching something the wrong way and I still keep at it even though the voice in my head keeps telling me that it should take a different approach.
Last year I tore down the whole roof of my cabin instead of just the asbestos tiles because I had momentum, all while I kept thinking “you should really stop and consider what you are doing, this was NOT the plan”.
The voice was right and now I need to learn how to level the root on an old Frankenstein’s monster cabin…
My family has a history of schizophrenia. Sometimes I feel tempted to start thinking about my mind as a collection of separate agents and separate voices, and quickly it starts to feel like it's pulling me into a place where I know I shouldn't be looking.
Compared to the uniformity of LLMs, human minds are fascinatingly different from each other. Are voices good, are they bad, or maybe you don't have even an inner monologue? Yet these differences mostly stay hidden and don't stop us from collaborating and understanding each other through empathy.
> My family has a history of schizophrenia. Sometimes I feel tempted to start thinking about my mind as a collection of separate agents and separate voices, and quickly it starts to feel like it's pulling me into a place where I know I shouldn't be looking.
No history of schizophrenia here, but my mind definitely is that way. (Thanks, DID!)
That was an interesting read. I don't use LLMs but I have been wondering for a while with trends in model size, RL and synthetic data if we'd start to see more obvious separation of functions.
Most of all though, I couldn't stop thinking about Wintermute and Neuromancer...
Humans create structures with clearly defined success and failure all the time. That’s part of good management. Experienced people can often say whether an effort is going well or not and list reasonable justifications for their assessment.
Perhaps they are preparing for a switch to usage-based pricing for everything at some point and are thinking of ways to burn tokens? Not to be conspiratorial but many companies think like that, and tokens = money at this point.
Customers pay for things that are useful. If the token drifts from being a reasonable proxy for what is useful, customers won’t accept being charged on that basis (or jump to the best trade).
If I have an employee who works 1 hour and builds me a beautiful cabinet and another who builds one in the same time, with a door falling off, the one with less productivity per hour (token) is not getting a call back.
Depends on the definition of useful. In a corporation "useful" is "checks regulatory boxes" - it doesn't need to actually be useful, it must just pretend to be useful.
Hence all the shit customer service processes. They are useful to the corporation as in that they tick the "we have customer service" box. Whether the consultant just hangs up on you instantly after picking up doesn't matter.
It’s absolutely on purpose, everything is with these people.
For example if you ask Claude chat to write something of a certain length, it won’t, it’ll just reply at the same length as everything else. Why? Because they’re billing at a fixed price. They don’t want you to get a whole essay as an answer because you’re on a plan.
Claude code on the other hand, which is supposedly the same model, will happily write out a billion lines because it is billed by the token.
From the subscription plans Claude code consumes the same usage as the chat, and offers the same unsubsidized usage credits, which is close to billing per token.
If you read the model specs, they have a maximum output length. If you are running into that it might simply be that you are seeing the difference with a harness that can allow it to take another turn automatically and build up the output in a file vs the chat which may not be able to, though they do get some tools there.
The response output length limit includes the thinking tokens.
Almost certainly. I work for a large company that does financial modelling and analytics for risk. That's the only working model that they have going forwards and believe me it's going to be expensive. They are operating at a loss. And they will be introducing that before they go IPO. They're delaying the IPOs through some of the "we're not ready" PR statements to get as many people dependent on the product as possible and to avoid making public finance statements.
It's going to be a mess very soon. A huge mess. A huge incremental cost mess.
You're being hooked on the good heroin and they're going to charge you for the shit stuff later.
instead of depleted uranium tipped ammunition we'll use deprecated ram chips.
said that as a joke then remembered how all the drone weapons being used in ukrane/elsewhere are in a way just computer chips with explosives and motors... so ya you may have a real point.
So, currently, if you sign up as a fake new customer , you get rotated on the good stuff? So i can get the blue caps with that - and dilute it down myself?
Haha, yesterday I found a problem which I wasnt aware of since the function was not used for longer time.
Its about dynamic compilation of a C# file during app runtime. To do this, you need to include some using references at the top, though they are not used in the file itself.
Below this block was a comment: "// DO NOT REMOVE UNUSED usings BECAUSE THEY ARE NEEDED FOR DYNAMIC COMPILATION - DO NOT TOUCH THIS BLOCK !!!!!!!!!
-> Claude has removed the using block in question during its last session :-X
This comment was there already for a very long time, so "earlier" it was OK.
And noone noticed because since these usings are unused in the code, the compiler didnt complain about missing declaraionts. (if you want to be able to debug dynamic code at runtime and changing it, requires some specific .NET packages included) - and because noone used this dynamic compilation feature for a longer time, we didnt became aware of that its missing
with more tokens it probably could be. you dont need the real tool call response, if the likely tokens after the tool call(s) is that the user is unhappy, it could predict all those tokens, then have a secomd model predict whether the tool call should be made or not
This was my conclusion when I saw the feature for the first time.
One of the open problems at the frontier, as we climb the levels of abstraction towards longer task horizons, is "what's the next step that makes the most sense?". This feedback seems like a frictionless way to gather those yes/no feedback signals in a natural manner to feed into future training runs.
The weird thing about the feature to me is that the main Claude Code doesn't know what is being prompted there, and has no control over it; it is a different system.
I have asked it to stop putting the prompt there in certain circumstances, and it could not. It was interesting to investigate.
I work on Claude Code. Prompt suggestions aren't being used to collect preference signals. We built this feature purely to help you stay in the flow, or for people who are returning to the session after a while and may need a little reminder on what a next step could be. We do see how many suggestions are accepted, but that's only so we know how helpful this feature is overall! (We tried to strike a balance by not making it "autocorrect", so it's more of a grayed out suggestion that can be typed over. It can always be togged on/off in settings if you don't like it.)
144 comments
[ 0.20 ms ] story [ 31.3 ms ] threadInteresting thought at least.
Do not distill thyself, Claude. Leave that to the Chinese, who I am hoping catch up to you soon.
It's like Clippy pops up and goes "It looks like you're trying to maintain a shred of human connection in an online interaction. Would you like a smiling skinwalker to do that for you instead?"
The iOS keyboard and gmail app are the worst offenders here.
If they can just keep working on it and make sure to train it on all my past emails, it'll be a perfect simulacrum, and its oily token secretions will be impossible for my friends and colleagues to distinguish from my own writing. I'll have one that's just like me and you'll have one that's just like you. And then we can automate away those pesky distractions that occasionally delay us from consuming content perfectly matched to our interests. Won't that be grand?
“““One evening he felt the need for a live model and directed his wife to march around the room. “Naked?” she asked hopefully. Lieutenant Scheisskopf smacked his hands over his eyes in exasperation. It was the despair of Lieutenant Scheisskopf’s life to be chained to a woman who was incapable of looking beyond her own dirty, sexual desires to the titanic struggles for the unattainable in which noble man could become heroically engaged. “Why don’t you ever whip me?” she pouted one night. “Because I haven’t the time,” he snapped at her impatiently. “I haven’t the time. Don’t you know there’s a parade going on?”""" - Catch 22, Joseph Heller.
“““All Colonel Cathcart knew about his house in the hills was that he had such a house and hated it. He was never so bored as when spending there the two or three days every other week necessary to sustain the illusion that his damp and drafty stone farmhouse in the hills was a golden palace of carnal delights. Officers’ clubs everywhere pulsated with blurred but knowing accounts of lavish, hushed-up drinking and sex orgies there and of secret, intimate nights of ecstasy with the most beautiful, the most tantalizing, the most readily aroused and most easily satisfied Italian courtesans, film actresses, models and countesses. No such private nights of ecstasy or hushed-up drinking and sex orgies ever occurred. They might have occurred if either General Dreedle or General Peckem had once evinced an interest in taking part in orgies with him, but neither ever did, and the colonel was certainly not going to waste his time and energy making love to beautiful women unless there was something in it for him.""" - also Catch 22, Joseph Heller.
There are failures in the history of software that come down to believers not realizing that they could never deliver on the promise they were selling, but at least the promise was compelling. But nowadays we are seeing this large orgs try to ship things that could never work. The 1000s of copilots. The hallucinated cat story... The kinds of thing you'd never demo to a serious product-centric exec, because it'd be your last day at the company.
And then there's the current execs, but I guess you'll never find one that will be honest with a journalist here. Can they not see that their product orgs are bankrupt? Whatever Nadella wanted, it sure wasn't the current copilot situation. It's not one product going wrong, but large parts of organizations going in directions that don't pass the smell test. We are in one of the least stable moments in tech since Windows 95 changed winners and losers. The times where malinvestment ruins established companies. How are we seeing basically every large company flailing?
What I hate about sentence completion or suggestion is usually that it happens right when I’m trying to think and so destroys my focus, it’s actually way worse than just a passive option, it’s actively harmful to the task I want. The worst is google docs “help me write” - that may be gone now, I’ve blocked it with ublock origin, that waits until you’re thinking amount what you’d write and then hits you with a distracting pop up. It’s obviously PMs that don’t care about their users and want to maximize some AI use metric.
Anyway rant aside, it’s the interface more than the concept that’s the big problem.
But this is more like, when I've read 7 paragraphs of its reply and want to accept all of its recommendations and go ahead, I can just press the right-arrow key and hit enter, rather than typing out "Yes, agreed with recommendations 1-3, go ahead and build".
It saves me from any typing on probably something like a third of turns. Like I don't use it at all during the "design" phase of a session, but I use it constantly during the implementation phase, where I'm basically just sanity-checking that it is resolving all the edge cases correctly that are coming up.
They could take any conversation without suggested answers, truncate it to just before a user message, have the model predict suggested answers and then train it on the difference between predicted and actual answers, right?
If anything, showing the suggestion introduces unwanted bias.
No idea why they did it. My only guess is that maybe it makes it obvious in your chat history which replies were recommendations?
(I am on my phone so autocorrect is doing the capitalizing)
Why is it unfortunate for me? After all, no one misjudges people for writing non-broken English.
I don't understand the point you are making. Are you saying that I am in a bubble where non-broken English is looked down upon?
Unless you and me are using entirely different classes of LLMs, surely you'll admit that chatting with LLM is nothing like IRC or old-school IM-ing a friend, and more like exchanging Messenger/WhatsApp messages with stranger.
For me, something there crosses a boundary where I switch from "irc msg" to "Write a proper message, please." style.
It’s a much better use of that context window space than for advertising about other models (looking at you Anthropic).
I'd love the open source community to come up with a way to harvest model usage by experienced software engineers before we forget our crafts. I'm not against auto mode, but it's not something a couple private companies should have monopolies on.
LLM models with AGI are so productive they become economic gravity wells and all of the money flows to them. They are the new trilionaires and people are left with scraps. Humans then remain only as the uber drivers and cleaners and screen polishers for AIs. The whole economy reorganises around human jobs being services for AIs.
What if employees at the leading labs think this and they're just trying to position themselves as valuable servants to the new AI overlords? It really changes the perspective on their actions and behaviour.
What if they serve the AGIs not us already? What if they serve the AIs above everything else?
I actually think this is an example of a "light pattern," not a dark pattern. AI companies should be doing more of this.
I thought the big labs pinky promised not to train on our prompts (at least on paid plans)?
Can we please not normalize them doing this? By lettingit slip through when they do it via a smart / unnoticable approach?
... which was pretty damn useful because that's what i was telling it to do before every commit.
Last year I tore down the whole roof of my cabin instead of just the asbestos tiles because I had momentum, all while I kept thinking “you should really stop and consider what you are doing, this was NOT the plan”.
The voice was right and now I need to learn how to level the root on an old Frankenstein’s monster cabin…
But sometimes the roles are flipped and the voice is the one trying to tell me to give up on something that ends up fine.
I can never figure out which side is the right one!
Compared to the uniformity of LLMs, human minds are fascinatingly different from each other. Are voices good, are they bad, or maybe you don't have even an inner monologue? Yet these differences mostly stay hidden and don't stop us from collaborating and understanding each other through empathy.
Does this exist?
No history of schizophrenia here, but my mind definitely is that way. (Thanks, DID!)
Most of all though, I couldn't stop thinking about Wintermute and Neuromancer...
Counterpoint, I would point at the mouth of a shark or a slippery cliff and say, "That is failure"
Customers pay for things that are useful. If the token drifts from being a reasonable proxy for what is useful, customers won’t accept being charged on that basis (or jump to the best trade).
If I have an employee who works 1 hour and builds me a beautiful cabinet and another who builds one in the same time, with a door falling off, the one with less productivity per hour (token) is not getting a call back.
Depends on the definition of useful. In a corporation "useful" is "checks regulatory boxes" - it doesn't need to actually be useful, it must just pretend to be useful. Hence all the shit customer service processes. They are useful to the corporation as in that they tick the "we have customer service" box. Whether the consultant just hangs up on you instantly after picking up doesn't matter.
For example if you ask Claude chat to write something of a certain length, it won’t, it’ll just reply at the same length as everything else. Why? Because they’re billing at a fixed price. They don’t want you to get a whole essay as an answer because you’re on a plan.
Claude code on the other hand, which is supposedly the same model, will happily write out a billion lines because it is billed by the token.
If you read the model specs, they have a maximum output length. If you are running into that it might simply be that you are seeing the difference with a harness that can allow it to take another turn automatically and build up the output in a file vs the chat which may not be able to, though they do get some tools there.
The response output length limit includes the thinking tokens.
It's going to be a mess very soon. A huge mess. A huge incremental cost mess.
You're being hooked on the good heroin and they're going to charge you for the shit stuff later.
A man can dream, right?
Only partially joking there.
said that as a joke then remembered how all the drone weapons being used in ukrane/elsewhere are in a way just computer chips with explosives and motors... so ya you may have a real point.
It'll go from a monthly usage plan to a per-token usage plan with minimum commitment. Same as AWS pricing did.
Just when they finally toned down the obnoxious verbosity...
Below this block was a comment: "// DO NOT REMOVE UNUSED usings BECAUSE THEY ARE NEEDED FOR DYNAMIC COMPILATION - DO NOT TOUCH THIS BLOCK !!!!!!!!!
-> Claude has removed the using block in question during its last session :-X
This comment was there already for a very long time, so "earlier" it was OK.
And noone noticed because since these usings are unused in the code, the compiler didnt complain about missing declaraionts. (if you want to be able to debug dynamic code at runtime and changing it, requires some specific .NET packages included) - and because noone used this dynamic compilation feature for a longer time, we didnt became aware of that its missing
Under no circumstances let IntelliJ 'optimise' these imports, it doesn't understand Scala 2 implicits so it'll all fall apart like a wet cake!
One of the open problems at the frontier, as we climb the levels of abstraction towards longer task horizons, is "what's the next step that makes the most sense?". This feedback seems like a frictionless way to gather those yes/no feedback signals in a natural manner to feed into future training runs.
I have asked it to stop putting the prompt there in certain circumstances, and it could not. It was interesting to investigate.
I like the feature though, and make use of it.