It's the error rate. That's what everyone found when they were trying to go Full Auto with OpenClaw in February.
You can rely on it like 95% of the time but that means if you keep it running continuously the error rate rapidly approaches 100%. That's getting a little better with each release, and it might actually hit the point where you can more or less trust it indefinitely (on well defined workflows).
Or at least it would, if context window permitted...
> The software world is very close to building a super intelligent senior software developer. Companies like this will ask all the best things a software engineer does automatically. Now claude will add it into the coding agents itself.
Except Claude is more expensive than an actual senior software developer. Otherwise, why are many companies terrified of the usage bill that gets printed on the invoice?
The nonsense in "tokenmaxxing" was a complete marketing scam and illusion of cheap tokens which in reality were heavily subsidized.
The entire point is detecting bad code before it reaches production. [0] AI generated or not.
I've been having a very nice time with Fable. I cooked up an Anki clone in like half an hour, with tech it's not familiar with. Nothing too ground breaking, but I was very pleased!
I think Opus might be on similar level for most of what I'm doing, but I haven't used it much recently, so I can't remember the difference. So I guess I'll find out on the 7th when they pull the plug again! (Free-ish trial of Fable ending.)
That being said, I tried using other frontier models to help with a Pong clone the other day and they were introducing new bugs at approximately the same rate as they were fixing it. On Pong!! I found that amusing because I couldn't think of a simpler game, so it didn't inspire confidence.
Fable's doing just fine on an online multiplayer game though. I have no idea how that works. (Maybe it would fail Pong too?? I haven't tested that!)
Isn’t this just a form of the bitter lesson? Our attempts to make engineered context and agents will simply be made obsolete with bigger and better models. Those transcripts are probably extremely useful for lesser capable models, and near unnecessary for frontier ones, maybe?
Yeah, the question is whether this applies to all of context management.
I've been using a custom harness based on https://minimal-agent.com/ (itself based on swe-mini-agent), which is like 50 lines for the core logic. Bash is all you need.
For small tasks, I find it's about 8x faster (and uses 8x fewer tokens) than the standard harness for each model.
For bigger tasks I haven't tested it much. It seems to work too but I think they're a bit less focused and productive in that case. It could be that those big harnesses' 20k token system prompts are doing something important with regard to steering software development workflows. (e.g. I heard Fable has a custom system prompt in Claude Code which might explain its markedly more proactive behavior.)
So I want to say there's still a lot of value in context engineering though it seems to diminish with each model release (since they're fine tuned on mostly non stupid behavior and need less hand holding).
> So I want to say there's still a lot of value in context engineering though it seems to diminish with each model release
I can't see how it would diminish unless you are literally working on public domain stuff. Unless stuffing context becomes cost effective and will not affect AI reasoning (this will be much harder), I don't see why context engineering is hear to stay until we have close to AGI.
I've wondered this. We have chain-of-thought, harnesses, etc. — workarounds of a sort due to lack of core model capabilities. But I am very curious if much better next token prediction would simply obsolete that whole setup or not. Either way, the answer would be very revealing.
interesting take. I think I disagree, but I like this take a lot and I had to think about it.
First, I think that models still need a context layer. One way to think about 'context' is as a form of compression. You provide the model context because it makes it easier for the model to figure out what to do. Even in a world with infinite model capacity and infinite model context, this is still useful because it allows the model to avoid rederiving everything from first principles every time. As long as models perform better using fewer tokens and as long as we care about token spend, context is a useful (necessary?) shortcut.
Once you bite that you need some form of context layer, the question is which. Here I do agree that it is better to work with what the models will find familiar (markdown files colocated with code, for eg). But this speaks to over-engineered solutions not understanding their main user (the agent) more than it does the need or lack there of.
A) Context and prompt cut the search space for next token generation. That’s pretty useful.
B) The other use of context is that it introduces entirely new information via RAG
B will never go away (as others pointed out). A, well that’s just something we’re all going to keep getting surprised at. We’ll barely give it any direction or context and the newer models will simply find the happy path.
The author is kind of suggesting that their context wasn’t really necessary to get the happy output, I think.
Chain of reasoning is a lot of context to guide token generation, but we simply see that newer models don’t need that context to get to the answer. I’m mostly reiterating this because there’s a hot take here, and that is this agentic stuff may be waived away by magic frontier-llm wand , all of a sudden.
They do, but that doesn’t necessarily mean it actually needed to for many questions. It’s on by default, in a way. You can probably prompt these models with “and don’t reason about it, just give me the answer” and probably get a comparably good response without it using reasoning tokens for many things.
I agree with the take not to bother with a sophisticated memory system. Anything worth remembering should be in docs, guides, source comments, commit messages or tickets. You don't need another layer, every conceivable granularity is already covered by existing best practices
Especially a layer that is largely out of band in a project (i.e. ~/.claude/…). In any project where I’ve needed memory I just add a line to AGENTS.md telling it to use MEMORY.md to save memories or STATUS.md to track progress.
I've been enjoying having a little todo file the agent updates as it goes along, because then I can keep track of progress without scrolling through aeons of "Combobulating..."
Also if context runs out you can just do "cat todo.md | agent" and you're off to the races again.
Yep all my projects start with a PLAN.md at the root, and that acts as the ‘save file’ recording our progress over time. My session always ends with updating the plan file with what’s been done, and the next session always begins, as you suggest, with consuming the current state of the plan doc.
I do think we need another layer, but it should be a routing layer. I am finalizing my pi-brains extension for Pi (https://github.com/earendil-works/pi) which does this:
Right now "humans" need to define the routing rules for how to access information, but I will support what I call "knowledge agents" that can monitor conversations to inject context when needed.
It looks like an interesting experiment. But a hard problem since it needs to store useful information and be able to inject it at the right time. It will also need to not be redundant to the information already stored.
What do you think is the potential value that you might get out of this, which is not already available with the existing options?
This is a hard problem, but one worth solving, I think, since it means less tokens and better AI reasoning. I believe LLMs are good enough that, if given the right context, it can very much solve almost all tasks.
If this works, it means we can probably get by with smaller models (since it doesn't need to know everything). LLMs are pattern matchers, and if you can provide them with the right shape (context), they should produce the expected output.
For my solution to work, you need business buy-in, which I don't think will be a problem. Enterprise wants to know how tokens are being spent, so I can see them wanting structured analysis during code reviews.
What may also not be obvious is that the information is ultimately designed to live with your code. Lessons and notes are designed to be mapped to files, so if you want to know why a piece of code is implemented in a certain way, you can have the LLM filter by files to help find the needle in the haystack.
It is a hard problem, but the only missing piece is discipline, which I believe business leaders will not have an issue with enforcing since we are ultimately talking about eliminating/significantly reducing the bus factor in our code.
There is some value to agents being able to query the history of work done, docs aren't a good place to accumulate negative evidence for example, but it can be tagged in traces so that it's efficient to look up as needed. Additionally, docs rot while traces can be tagged with commit hashes and other things that make their lifetime clearer.
> I agree with the take not to bother with a sophisticated memory system. Anything worth remembering should be in docs, guides, source comments, commit messages or tickets. You don't need another layer
That is a sophisticated memory system though -- maybe not to you experienced humans!
> I believed this so strongly that my company built an entire product around this concept. I used to tell folks that "session transcripts were the new oil," that they were more valuable than the code itself.
> […]
> We don't really write code by hand anymore.
Honestly, isn't this just influencer spam? What possible value is there in reading about people who used to have products, but no longer write their own code, complaining about the inscrutable prediction machine they have handed that job and their livelihoods to?
Like, if you have complaints about the thing, perhaps you should address them to your supplier directly. None of your readers can help, and nobody's magic folk solution to your problem is better than yours.
And there are so many of these sorts of posts. Are we not entirely cooked?
I am now in the process of fixing code I wrote using AI. I have come to the realization that AI can't really write software and I am annoyed that it took me that long (months) to realize that.
This is quite terrifying to me, because I have a feeling I will soon come to the same conclusion.
I’m starting to see some really glaring omissions in code I’m responsible for (using Opus) that at first (and second) look seemed fine, but really isn’t.
I talked with a friend on a different field (academic) and he had to re-review all things written by AI. Basically, he used AI to read/summarize/find stuff in large academic papers but realized later that many times AI makes glaring mistakes that on a first read pass the smell test.
From my perspective it felt like understanding that the machine has no desires helped refine my usage.
I can ask it to be curious, and it will reply with what people think curiosity should look like, but it’s a simulation of an emotion it will never be driven by.
The ramifications become apparent when you engage in activity like cross-domain discovery.
I am more fully invested in finding out ways AI can support me (documentation, code analysis, bughunting), though my experience with Claude as a bughunter is that it can miss the absolutely obvious if it is not in the shape it is expecting.
More generally I am interested in burnout-avoidance tools; things that help me start, finish, things that write tests I guess, certainly code scaffolding.
But I am fully unconvinced that my burnout will be improved by ending up owning the responsibility for wobbly or inscrutable AI-generated code with potential landmines in it; that will keep me up at night just the same.
I don't understand this line or reasoning. People use various cryptocurrencies to buy and sell legitimate products and services every day. Is the argument just that they could probably have done it some other way?
People do, but I personally don't know anyone who does. And I don't exactly live in a bubble, half of my friends were into crypto at one point or the other.
Look man, I’ve got a MMO that I’m working on that’s set in 2014 where everyone is a programmer in SV. It’s a period piece. I NEED as much blog training data of this type so that my NPCs can talk in a historically accurate way (god bless Medium.com, a historical treasure trove of a bygone medieval era).
It’s gonna be a living breathing world, you see. You’re going to be like “omg, this game even accurately captured the blog posts, woah”.
Something about this idea really resonates with certain personality types. I equate it to the Zettelkasten hype phase from several years ago. People (...like me..) got really wrapped up in the belief that the process was more important that the content. "Linking" was an "activity." Something good will happen as long as you (a) take notes on stuff and (b) link them to other notes on stuff.
You see the same thing with the session transcripts people. They're building ever more sophisticated setups of indexing and storing and cross referencing every conversation they've ever had on the (I would argue) mistaken belief that the transcripts are the valuable part, rather than the uncomfortable part where you go do something. A lot of it, I say from falling in the trap, is fancy procrastination.
(Although, I have found myself jealous on many occasions where their fancy system retrieves something they vaguely recall from a conversation they had 3 months ago. So, who knows.)
> Something about this idea really resonates with certain personality types.
Like ancient people? Because "new oil" whilst I get what it might imply sounds bad to me. Oil has been superseded in many places so "new oil" is like going backwards still.
Reference: data is the new oil is a term coined in 2006.
Absolutely agreed. Anyone who's a serious procrastinator sooner or later noticed that pattern of theirs in which they spent immense effort on optimizing the process instead focusing on the outcome they really wish — just don't really believe they can deliver it.
Occasionally posts like this do get the attention of the company responsible, more than an email does... but indeed that's like a one in a million situation
> Like, if you have complaints about the thing, perhaps you should address them to your supplier directly. None of your readers can help, and nobody's magic folk solution to your problem is better than yours.
I think you may just misunderstand the point of having / writing a personal blog. I write because it's fun! Whether the reader gets any value out of reading it is almost entirely beside the point.
(Also several comments here directly post a fix to the problem stated in the blog post, so readers can and do often help)
> I believed this so strongly that my company built an entire product around this concept. I used to tell folks that "session transcripts were the new oil," that they were more valuable than the code itself.
This is pretty funny because it's about the depth of understanding of every 'AI expert' on Linkedin. People who praise the context window as basically magic have no idea how any of this works.
Strongly agree here. claude-code’s memory system is occasionally useful but much more often harmful, pulling in obsolete info that muddies the waters about current tasks. I’ve frequently seen Claude’s own memories severely mislead it.
My guess is that has something to do with the training process leaving models unable to differentiate between “what’s happening now” and “what happened before”. Perhaps if making inferences from memories was actually part of the training process things would be different but my sense is that as an inference-time-only feature this just gets the models confused.
Humans make memories constantly, but they also forget things that are no longer relevant. Until Claude can do that, it means the LLM will have an ever-increasing, ever more fragmented context.
And LLMs are NOT intelligent enough to survive even mild context poisoning.
When claude goes down a wrong path, I tend to clear context and write a new prompt that helps guide it down the correct path.
Whatever thinking or context that led it there has inertia and tends to be sticky, otherwise.
Pretty annoying when it brings those up again later from memory...
t once had to tell claude 3-4 times to stop assuming the state of a system was the way it kept iterating it was cause it was in it's memory. I repeatably told it to otherwise and it just never updated it's memory and instead kept referencing it's memory about the state of a particular system
In my harness I have all the code auto injected at startup (doing mostly very small codebases).
I found that every model will still manually check every file/function, they immediately assume that anything in context is stale.
That's sensible because often the user edits stuff while they're running.
What it does is save it from having to grep blindly about the codebase. But I think I'd get roughly the same benefit by just dumping the function headers then.
Even with memory off this occurs within a conversation.
It is like an annoying friend, who remembers something from a past conversation, that you have grown and developed from, but they still want to hold it against you.
It's because it mostly doesn't matter what you are trying to get the code to do. What matters is what the code does.
Session logs can absolutely be useful, but not when building further. It's just that that the place they slot in is during validation. You know, that place between the markdown plan and CI passing, where there's 800 new lines of code and it all seems sort of fine when you click around?
Session logs can show you what sort of manual validation happened. CI will run the tests you had, and the code will show you what new unit tests were added, but session logs can show you that the agent drove the app with Playwright, or that the agent read and considered the prod config as well as the dev config.
Nothing bulletproof, but not every piece of validation work merits a test in the repo that lives forever. We've gotten a lot of mileage out of re-analyzing the sessions, figuring out where the agent made decisions without asking, and forcing the agent to consider validation for those decisions. That's the sort of thing that's hard to dictate up front but easy to highlight with the session logs.
I specifically disabled claude memory in a project because it kept writing down thigns to memory that didn't need to be in memory, including severly wrong statements that then would confuse it later. At some point it got re-enabled automatically which had me ask claude itself to "turn it the fuck off" by which it promptly figured out that both ("autoMemoryEnabled": false, "autoDreamEnabled": false) are necessary and need to be at the user home settings, not in a project override (which is what I had with the original setup that eventually got ignored by a CC update).
I agree with other commenters here, if anything is worth being rememebered, it will be in code comments, git commit messages, CLAUDE.md or other formal documentation. The auto memory system just causes confusion and leaves stale and outdated information written down.
Its an interesting thought experiment as well, I originally thought that having the model write down memory files by itself would be a nice addition, but after playing around with it, it became clear to me that good as an idea turns out bad in practice because the model can't correctly gauge what deserves being stored as a memory.
> I believed this so strongly that my company built an entire product around this concept. I used to tell folks that "session transcripts were the new oil," that they were more valuable than the code itself.
This is infuriatingly common wrt talking/writing about how to use AI effectively. All of the "this is how you write an AGENTS.md" and "you need to talk to it like X to optimize it". Like sure, you can believe that as much as you want but unless you provide some evidence you can keep your shitty CLAUDE.md to yourself and don't pollute the whole company's git repo, thanks.
I found that if you allow any low value things into memory, Claude will notice that established pattern and start trying to add low value memories at an ever increasing pace.
>We have found zero performance benefit on SWE tasks when agents have search access to their previous transcript sessions
I refuse to believe this is true. The ability for an agent to find information from before a compaction is incredibly useful. At compaction time it's impossible to know what exactly may be still needed.
With the million-context-window models we never hit compaction, observed over hundreds of sessions. What are you doing that has you hitting compaction regularly?
For me logs can chew through a lot of tokens. And when the agent is trying a bunch of different experiments and then it may need to refer to what happened previously.
Million context models also are still not effective for the entire context size.
I like the memory system, in general. For reference I'm using mostly Opus 4.8 + Max effort. It will often pull things out of memory that are relevant. Like I'll ask it to come up with a few options I should consider for, say, a self-hosted OIDC provider and it'll say things like "Considering the size of your operations team, this might be a better fit because of X and Y".
Now, I'll agree that this is probably the sort of thing I should put in the CLAUDE.md, but in this case it wasn't on my radar to put that in my CLAUDE.md, so it was nice that it surfaced that.
It does sometimes go awry though. Today I was asking about a problem I was having authenticating, and it said "you may be running into this trusted proxy setting because you put your apps behind an haproxy". That is true of 95% of our apps, so it was worth mentioning, but in this case it was not so I had to correct it. But, I'm glad it mentioned it because if we did have it proxied it could have saved me a lot of time.
It seems like a prerequisite is a certain level of world model and associated reasoning ability. Your examples are entirely dependent on the past context being relevant to the current situation. That's particularly tricky if you regularly ask about hypotheticals or problems that you're assisting someone else with. A human would probably ask clarifying questions such as "is this for the operations team at X? are they still size Y?" and "is this app proxied like the others you mentioned in the past?" rather than assuming.
There's also a noticable hierarchy to such context that needs to be correctly modeled - you could for example be involved with multiple teams of different sizes that are subject to different rules which is something a human would understand naturally.
You say that, and that's a possibility, and maybe in your experience that is true. But I'd say I've found actual humans will frequently use existing internal models of something unless they have a reason to believe something has changed to invalidate that. We even have a well worn joke about what happens when you assume.
>you could for example be involved with multiple teams of different sizes that are subject to different rules
Just to clarify, my experience with the memory of these tools doesn't provide any data (for or against) how well/poorly the memory may work with someone who is involved with multiple teams.
> I'd say I've found actual humans will frequently use existing internal models of something
Sure, I didn't mean to claim that humans perform perfectly. Merely to reference a particular approach in order to illustrate my point about modeling concepts.
> doesn't provide any data (for or against)
It seems a foregone conclusion that it would confuse details between the different contexts in a way that a human (ie an entity with a certain level of world model and reasoning ability) would not be expected to.
I remember when OpenAI announced ChatGPT now will remember stuff between sessions. Oh, you mean find random trivia about me and copy paste it between prompts without out my explicit consent.
”compare these three cars. Oh btw I am a data engineer, and my moms maiden name is Joana, and I am allergic to bad poetry. And code should be DRY, I prefer SQL over Python and what’s the most poisonous flower in Scandinavia?”.
I’ve had so much wierd output because context is ”””memorized””” and bleeding into completely unrelated projects and conversations. It’s the first feature I turn off.
I have to imagine these types of features are for people who treat chatgpt like their friend/therapist/girlfriend/assistant/... instead of people who use it to answer questions
I mean, it’s pretty clear the people who work on Claude Code aren’t actually looking at what they’re implementing. The thought behind this feature seems like it goes nowhere beyond “oh wouldn’t it be nice if Claude could remember things about you? Ok Claude go implement this” and nobody bothered to see if it was useful or helpful.
Its certainly true at the moment, but give it 10 years and we might have systems that are much cheaper and much better at context management than they are now.
(Apologies to anyone who is under the impression that we were very likely going to be at the singularity in 10 years time. Possible != very likely)
Sure, but it’s equally likely that we hit a point where scaling becomes economically unviable because we can’t come up with enough algorithmic improvements to break free of the tyranny of log linear scaling. (I’m not sure how many 2x in token cost people would be willing to pay)
100 comments
[ 814 ms ] story [ 2644 ms ] threadToggle it off and never think about it again.
The software world is very close to building a super intelligent senior software developer.
Damn, I didn't see this coming.
Its first the build the intelligent builder. We will figure out what we want to build later.
Yeah. Two more weeks, as they say. Just need to iron out some kinks.
You can rely on it like 95% of the time but that means if you keep it running continuously the error rate rapidly approaches 100%. That's getting a little better with each release, and it might actually hit the point where you can more or less trust it indefinitely (on well defined workflows).
Or at least it would, if context window permitted...
Except Claude is more expensive than an actual senior software developer. Otherwise, why are many companies terrified of the usage bill that gets printed on the invoice?
The nonsense in "tokenmaxxing" was a complete marketing scam and illusion of cheap tokens which in reality were heavily subsidized.
The entire point is detecting bad code before it reaches production. [0] AI generated or not.
[0] https://sketch.dev/blog/our-first-outage-from-llm-written-co...
I think Opus might be on similar level for most of what I'm doing, but I haven't used it much recently, so I can't remember the difference. So I guess I'll find out on the 7th when they pull the plug again! (Free-ish trial of Fable ending.)
That being said, I tried using other frontier models to help with a Pong clone the other day and they were introducing new bugs at approximately the same rate as they were fixing it. On Pong!! I found that amusing because I couldn't think of a simpler game, so it didn't inspire confidence.
Fable's doing just fine on an online multiplayer game though. I have no idea how that works. (Maybe it would fail Pong too?? I haven't tested that!)
I've been using a custom harness based on https://minimal-agent.com/ (itself based on swe-mini-agent), which is like 50 lines for the core logic. Bash is all you need.
For small tasks, I find it's about 8x faster (and uses 8x fewer tokens) than the standard harness for each model.
For bigger tasks I haven't tested it much. It seems to work too but I think they're a bit less focused and productive in that case. It could be that those big harnesses' 20k token system prompts are doing something important with regard to steering software development workflows. (e.g. I heard Fable has a custom system prompt in Claude Code which might explain its markedly more proactive behavior.)
So I want to say there's still a lot of value in context engineering though it seems to diminish with each model release (since they're fine tuned on mostly non stupid behavior and need less hand holding).
I can't see how it would diminish unless you are literally working on public domain stuff. Unless stuffing context becomes cost effective and will not affect AI reasoning (this will be much harder), I don't see why context engineering is hear to stay until we have close to AGI.
Bear in mind that brain architecture is learnt too - just over a much longer timescale than an individual lifetime.
First, I think that models still need a context layer. One way to think about 'context' is as a form of compression. You provide the model context because it makes it easier for the model to figure out what to do. Even in a world with infinite model capacity and infinite model context, this is still useful because it allows the model to avoid rederiving everything from first principles every time. As long as models perform better using fewer tokens and as long as we care about token spend, context is a useful (necessary?) shortcut.
Once you bite that you need some form of context layer, the question is which. Here I do agree that it is better to work with what the models will find familiar (markdown files colocated with code, for eg). But this speaks to over-engineered solutions not understanding their main user (the agent) more than it does the need or lack there of.
B) The other use of context is that it introduces entirely new information via RAG
B will never go away (as others pointed out). A, well that’s just something we’re all going to keep getting surprised at. We’ll barely give it any direction or context and the newer models will simply find the happy path.
The author is kind of suggesting that their context wasn’t really necessary to get the happy output, I think.
Chain of reasoning is a lot of context to guide token generation, but we simply see that newer models don’t need that context to get to the answer. I’m mostly reiterating this because there’s a hot take here, and that is this agentic stuff may be waived away by magic frontier-llm wand , all of a sudden.
I thought each new generation typically used more reasoning tokens?
Also if context runs out you can just do "cat todo.md | agent" and you're off to the races again.
I do think we need another layer, but it should be a routing layer. I am finalizing my pi-brains extension for Pi (https://github.com/earendil-works/pi) which does this:
https://github.com/gitsense/pi-brains
Right now "humans" need to define the routing rules for how to access information, but I will support what I call "knowledge agents" that can monitor conversations to inject context when needed.
What do you think is the potential value that you might get out of this, which is not already available with the existing options?
If this works, it means we can probably get by with smaller models (since it doesn't need to know everything). LLMs are pattern matchers, and if you can provide them with the right shape (context), they should produce the expected output.
For my solution to work, you need business buy-in, which I don't think will be a problem. Enterprise wants to know how tokens are being spent, so I can see them wanting structured analysis during code reviews.
What may also not be obvious is that the information is ultimately designed to live with your code. Lessons and notes are designed to be mapped to files, so if you want to know why a piece of code is implemented in a certain way, you can have the LLM filter by files to help find the needle in the haystack.
It is a hard problem, but the only missing piece is discipline, which I believe business leaders will not have an issue with enforcing since we are ultimately talking about eliminating/significantly reducing the bus factor in our code.
If you look at https://github.com/gitsense/smart-ripgrep, you can get a better sense of how context can be injected when it is needed.
That is a sophisticated memory system though -- maybe not to you experienced humans!
> I believed this so strongly that my company built an entire product around this concept. I used to tell folks that "session transcripts were the new oil," that they were more valuable than the code itself.
> […]
> We don't really write code by hand anymore.
Honestly, isn't this just influencer spam? What possible value is there in reading about people who used to have products, but no longer write their own code, complaining about the inscrutable prediction machine they have handed that job and their livelihoods to?
Like, if you have complaints about the thing, perhaps you should address them to your supplier directly. None of your readers can help, and nobody's magic folk solution to your problem is better than yours.
And there are so many of these sorts of posts. Are we not entirely cooked?
I can ask it to be curious, and it will reply with what people think curiosity should look like, but it’s a simulation of an emotion it will never be driven by.
The ramifications become apparent when you engage in activity like cross-domain discovery.
More generally I am interested in burnout-avoidance tools; things that help me start, finish, things that write tests I guess, certainly code scaffolding.
But I am fully unconvinced that my burnout will be improved by ending up owning the responsibility for wobbly or inscrutable AI-generated code with potential landmines in it; that will keep me up at night just the same.
I do personally know people who pay for regular products with cryptocurrency. Including myself.
It’s gonna be a living breathing world, you see. You’re going to be like “omg, this game even accurately captured the blog posts, woah”.
Something about this idea really resonates with certain personality types. I equate it to the Zettelkasten hype phase from several years ago. People (...like me..) got really wrapped up in the belief that the process was more important that the content. "Linking" was an "activity." Something good will happen as long as you (a) take notes on stuff and (b) link them to other notes on stuff.
You see the same thing with the session transcripts people. They're building ever more sophisticated setups of indexing and storing and cross referencing every conversation they've ever had on the (I would argue) mistaken belief that the transcripts are the valuable part, rather than the uncomfortable part where you go do something. A lot of it, I say from falling in the trap, is fancy procrastination.
(Although, I have found myself jealous on many occasions where their fancy system retrieves something they vaguely recall from a conversation they had 3 months ago. So, who knows.)
Like ancient people? Because "new oil" whilst I get what it might imply sounds bad to me. Oil has been superseded in many places so "new oil" is like going backwards still.
Reference: data is the new oil is a term coined in 2006.
We're in 2026. See what I mean.
I think you may just misunderstand the point of having / writing a personal blog. I write because it's fun! Whether the reader gets any value out of reading it is almost entirely beside the point.
(Also several comments here directly post a fix to the problem stated in the blog post, so readers can and do often help)
I used to blog, as it goes, and I have supported and enabled many more, so no, not really.
This is pretty funny because it's about the depth of understanding of every 'AI expert' on Linkedin. People who praise the context window as basically magic have no idea how any of this works.
"Spicy Autocomplete", I've heard it called.
My guess is that has something to do with the training process leaving models unable to differentiate between “what’s happening now” and “what happened before”. Perhaps if making inferences from memories was actually part of the training process things would be different but my sense is that as an inference-time-only feature this just gets the models confused.
And LLMs are NOT intelligent enough to survive even mild context poisoning.
Pretty annoying when it brings those up again later from memory...
> Don't start generating an auto-memory entry before asking me. Ask first, write only if I confirm — no speculative drafting.
No more crap after this.
Incidentally I don’t recall Opus 4.8 asking me once in the past few weeks. Older models did ask semi-frequently.
I found that every model will still manually check every file/function, they immediately assume that anything in context is stale.
That's sensible because often the user edits stuff while they're running.
What it does is save it from having to grep blindly about the codebase. But I think I'd get roughly the same benefit by just dumping the function headers then.
It is like an annoying friend, who remembers something from a past conversation, that you have grown and developed from, but they still want to hold it against you.
Session logs can absolutely be useful, but not when building further. It's just that that the place they slot in is during validation. You know, that place between the markdown plan and CI passing, where there's 800 new lines of code and it all seems sort of fine when you click around?
Session logs can show you what sort of manual validation happened. CI will run the tests you had, and the code will show you what new unit tests were added, but session logs can show you that the agent drove the app with Playwright, or that the agent read and considered the prod config as well as the dev config.
Nothing bulletproof, but not every piece of validation work merits a test in the repo that lives forever. We've gotten a lot of mileage out of re-analyzing the sessions, figuring out where the agent made decisions without asking, and forcing the agent to consider validation for those decisions. That's the sort of thing that's hard to dictate up front but easy to highlight with the session logs.
I agree with other commenters here, if anything is worth being rememebered, it will be in code comments, git commit messages, CLAUDE.md or other formal documentation. The auto memory system just causes confusion and leaves stale and outdated information written down.
Its an interesting thought experiment as well, I originally thought that having the model write down memory files by itself would be a nice addition, but after playing around with it, it became clear to me that good as an idea turns out bad in practice because the model can't correctly gauge what deserves being stored as a memory.
This is infuriatingly common wrt talking/writing about how to use AI effectively. All of the "this is how you write an AGENTS.md" and "you need to talk to it like X to optimize it". Like sure, you can believe that as much as you want but unless you provide some evidence you can keep your shitty CLAUDE.md to yourself and don't pollute the whole company's git repo, thanks.
I refuse to believe this is true. The ability for an agent to find information from before a compaction is incredibly useful. At compaction time it's impossible to know what exactly may be still needed.
Million context models also are still not effective for the entire context size.
Now, I'll agree that this is probably the sort of thing I should put in the CLAUDE.md, but in this case it wasn't on my radar to put that in my CLAUDE.md, so it was nice that it surfaced that.
It does sometimes go awry though. Today I was asking about a problem I was having authenticating, and it said "you may be running into this trusted proxy setting because you put your apps behind an haproxy". That is true of 95% of our apps, so it was worth mentioning, but in this case it was not so I had to correct it. But, I'm glad it mentioned it because if we did have it proxied it could have saved me a lot of time.
There's also a noticable hierarchy to such context that needs to be correctly modeled - you could for example be involved with multiple teams of different sizes that are subject to different rules which is something a human would understand naturally.
You say that, and that's a possibility, and maybe in your experience that is true. But I'd say I've found actual humans will frequently use existing internal models of something unless they have a reason to believe something has changed to invalidate that. We even have a well worn joke about what happens when you assume.
>you could for example be involved with multiple teams of different sizes that are subject to different rules
Just to clarify, my experience with the memory of these tools doesn't provide any data (for or against) how well/poorly the memory may work with someone who is involved with multiple teams.
Sure, I didn't mean to claim that humans perform perfectly. Merely to reference a particular approach in order to illustrate my point about modeling concepts.
> doesn't provide any data (for or against)
It seems a foregone conclusion that it would confuse details between the different contexts in a way that a human (ie an entity with a certain level of world model and reasoning ability) would not be expected to.
”compare these three cars. Oh btw I am a data engineer, and my moms maiden name is Joana, and I am allergic to bad poetry. And code should be DRY, I prefer SQL over Python and what’s the most poisonous flower in Scandinavia?”.
I’ve had so much wierd output because context is ”””memorized””” and bleeding into completely unrelated projects and conversations. It’s the first feature I turn off.
It'll assume I own a datacenter and have lots of gpus just because I asked to research things.
Its certainly true at the moment, but give it 10 years and we might have systems that are much cheaper and much better at context management than they are now.
(Apologies to anyone who is under the impression that we were very likely going to be at the singularity in 10 years time. Possible != very likely)
is the conclusion really that its just more important to create proper artifacts from any tricks that got the llm to understand the code better?
is the tool for searching the history just bad?