Love the idea, I was thinking about this problem a few weeks ago, but never got anywhere with it. I'm intrigued by the latency savings in particular, that sounds great.
One concern I have is that right now each session gets fresh "eyes" on the problem. Right now I find I get a lot of mileage out of a combination of long-running sessions and fresh ones. I worry with a single generated concept graph that gets only incremental refreshes will become stale slowly, and in subtle ways that are hard to detect. That could lead to semantic drift in the graph from reality, and every new session will take the drifted form as gospel. Have you run any long tests (weeks or longer) on this to make sure that this doesn't happen? My understanding is SWE Bench is only a point-in-time evaluation.
Also the graph is stored in the repo, right? How mergeable is it? I know I wouldn't want to do conflict resolution on that myself, and even Opus struggles to keep all the references correct (especially when comments are involved) when there's an B->C, A->B symbol rename.
Opus 5 specifically seems to be affected by a severe case of turbo-encabulationitis, almost as if it was trained to spit out as complex of sentences as it can.
And may your deity of choice help you if you decide to venture on subjects which you don't have a deep understanding of, as half the sentences it generates will be (barely) cohesive.
> almost as if it was trained to spit out as complex of sentences as it can.
It probably was, at least inadvertently.
Very easy to imagine that one of the post-training step (RLHF, DPO etc) reinforced the "sounds clever" behaviour.
Every task, your coding agent starts blind. Before it changes anything, it re-explores the repo: grep a term, open a file, follow an import, back out, try again. It is rebuilding a picture of a codebase it mapped an hour ago and threw away.
That rediscovery burns most of a run's tool calls, tokens, and latency, and it is pure overhead"
The author of this article brings up a very interesting problem -- that, at least as far as using LLM's as coders/coding assistants go, eventually context runs out and context related to the underlying codebase does too. This in turn burns tokens and in turn, wastes energy resources.
Historically (well, in the past couple of years!), a bunch of solutions have been proposed to address this problem (i.e., take abstracts/subsets/maps of code, write them to different databases and persistent storage methods, bring them back in when the LLM requires it, etc., etc.)...
But there's no really good solution to this problem (although, arguably Graft goes a lot farther than past tools and should be commended for that!) because the problem seems to lie in separate parts, across several problem domains:
1) LLM context window size -- limited. Anything that future LLM's do to make context windows larger will help ameliorate this problem.
2) Lack of a good way to represent a codebase to an LLM for training other than text.
In other words, first we need some kind of way to map codebases into Tensors rather than text (i.e., a higher-level "map" of the code) then train future LLM's on those code-specific Tensors.
3) Arguably, programming languages themselves share some of the blame...
Programming languages have historically been written so that an arbitrary corpus of text represents and can be interpreted and/or compiled into a computer program.
That is, while tools for mapping codebases exist, tools for directly training LLM's on those specific created "code maps" as Tensors, do not, do not seem to, or at least I'm currently unaware of any!
(Anyway, just thinking aloud...)
Graft looks good, and looks like it has made some serious inroads to solving the problem...
We should have moved past storing code in files, using the filesystem as the symbol database of programs, a long time ago. There was a lot of interesting research toward this in Haskell, for example, but also Academia in general. There’d be lots of value in using things like SQLite for example, or just ecosystem-specific containers that know about the layout and can present it to an IDE or LLM or a runtime in whatever shape is best suited to the task.
Love the idea, I was thinking about this problem a few weeks ago, but never got anywhere with it. I'm intrigued by the latency savings in particular, that sounds great.
One concern I have is that right now each session gets fresh "eyes" on the problem. Right now I find I get a lot of mileage out of a combination of long-running sessions and fresh ones. I worry with a single generated concept graph that gets only incremental refreshes will become stale slowly, and in subtle ways that are hard to detect. That could lead to semantic drift in the graph from reality, and every new session will take the drifted form as gospel. Have you run any long tests (weeks or longer) on this to make sure that this doesn't happen? My understanding is SWE Bench is only a point-in-time evaluation.
Also the graph is stored in the repo, right? How mergeable is it? I know I wouldn't want to do conflict resolution on that myself, and even Opus struggles to keep all the references correct (especially when comments are involved) when there's an B->C, A->B symbol rename.
I'm glad Claude is so recognisable, it lets me bounce right off the empty calorie language very efficiently.
Maybe this thing is great, but it cannot be determined with this presentation.
Opus 5 specifically seems to be affected by a severe case of turbo-encabulationitis, almost as if it was trained to spit out as complex of sentences as it can.
And may your deity of choice help you if you decide to venture on subjects which you don't have a deep understanding of, as half the sentences it generates will be (barely) cohesive.
> almost as if it was trained to spit out as complex of sentences as it can.
It probably was, at least inadvertently. Very easy to imagine that one of the post-training step (RLHF, DPO etc) reinforced the "sounds clever" behaviour.
"It tested well with the focus group..."
You can actually tell from the very first real sentence in the README:
> Efficiency is a 162-run controlled benchmark (same agent, same file tools, only the context differs).
It's so bad that I can tell it's Claude, specifically Opus 4.8/5.0, from the first 6 words.
Opus 5 is very paranoid on giving proofs for every thing it claimed. so I let it keep this one line.
Why these animations in the github README? why? It just made understanding anything more difficult
Does Claude's grep still prepend the relative path of the file before _every_ single line? Because that is nasty, especially in java projects.
Relative path? You sweet summer child. It prepends the absolute path for me all the time.
What's the benchmark against graphify?
Is this still cheaper when stale?
we use claude hooks. so that, post edits it auto updates and before using the graft tools also we check if it's already in sync.
Hooks also solve for the issue of the LLMs not calling our CLI tools instead of Grep.
I'm thinking more in terms of the case where my coworkers' agents aren't working via graft, and changes are made without the post-edit update hooks.
if the changes are on your local when you pull those changes from the remote, graft will be auto-synced before graft tool use.
Oh, of course. Brilliant.
>"The problem:
Every task, your coding agent starts blind. Before it changes anything, it re-explores the repo: grep a term, open a file, follow an import, back out, try again. It is rebuilding a picture of a codebase it mapped an hour ago and threw away.
That rediscovery burns most of a run's tool calls, tokens, and latency, and it is pure overhead"
The author of this article brings up a very interesting problem -- that, at least as far as using LLM's as coders/coding assistants go, eventually context runs out and context related to the underlying codebase does too. This in turn burns tokens and in turn, wastes energy resources.
Historically (well, in the past couple of years!), a bunch of solutions have been proposed to address this problem (i.e., take abstracts/subsets/maps of code, write them to different databases and persistent storage methods, bring them back in when the LLM requires it, etc., etc.)...
But there's no really good solution to this problem (although, arguably Graft goes a lot farther than past tools and should be commended for that!) because the problem seems to lie in separate parts, across several problem domains:
1) LLM context window size -- limited. Anything that future LLM's do to make context windows larger will help ameliorate this problem.
2) Lack of a good way to represent a codebase to an LLM for training other than text.
In other words, first we need some kind of way to map codebases into Tensors rather than text (i.e., a higher-level "map" of the code) then train future LLM's on those code-specific Tensors.
3) Arguably, programming languages themselves share some of the blame...
Programming languages have historically been written so that an arbitrary corpus of text represents and can be interpreted and/or compiled into a computer program.
That is, while tools for mapping codebases exist, tools for directly training LLM's on those specific created "code maps" as Tensors, do not, do not seem to, or at least I'm currently unaware of any!
(Anyway, just thinking aloud...)
Graft looks good, and looks like it has made some serious inroads to solving the problem...
We should have moved past storing code in files, using the filesystem as the symbol database of programs, a long time ago. There was a lot of interesting research toward this in Haskell, for example, but also Academia in general. There’d be lots of value in using things like SQLite for example, or just ecosystem-specific containers that know about the layout and can present it to an IDE or LLM or a runtime in whatever shape is best suited to the task.