If the result is always provably correct it doesn't matter whether or not it's different at the code level. People interested in systems like this believe that the outcome of what the code does is infinity more important than the code itself.
Since nobody involved actually cares whether the code works or not, it doesn't matter whether it's a different wrong thing each time.
If anyone cared enough they could look at the code and see the problem immediately and with little effort, but we're encouraging a world where no one cares enough to put even that baseline effort because *gestures at* the tests are passing. Who cares how wrong the code is and in what ways if all the lights are green?
Except one shoe is made by children in a fire-trap sweatshop with no breaks, and the other was made by a well paid adult in good working conditions.
The ends don’t justify the means. The process of making impacts the output in ways that are subtle and important, but even holding the output as a fixed thing - the process of making still matters, at least to the people making it.
And guess how much shoe companies make who manufacture shoes in sweatshop conditions versus the ones who make artisanal handcrafted shoes?
Btw in my metaphor, we - the programmers - are the kids in the sweatshop.
Even on the BigTech side being able to reverse a btree on the whiteboard and having on your resume that you were a mid level developer isn’t enough either anymore
If you look at the comp on that side, it’s also stagnated for decade. AI has just accelerated that trend.
While my job has been at various percentages to produce code for 30 years, it’s been well over a decade since I had to sell myself on “I codez real gud”. I sell myself as a “software engineer” who can go from ambiguous business and technical requirements, deal with politics, XYProblems, etc
That’s exactly my point. “Programming” was clearly becoming commoditized a decade ago.
But while you are clutching your pearls, where do you think your computer, clothes etc are being made?
Out of bounds behavior is sometimes a known unknown, but in the era of generated code is exclusively unknown unknowns.
Good luck speccing out all the unanticipated side effects and undefined behaviors. Perhaps you can prompt the agent in a loop a bnumber of times but it's hard to believe that the brute-force throw-more-tokens-at-it approach has the same level of return as a more attentive audit by human eyeballs.
I don’t review every line of code by everyone whose output I’m responsible for, I ask them to explain how they did things and care about their testing, the functional and non functional requirements and hotspots like concurrency, data access patterns, architectural issues etc.
For instance, I haven’t done web development since 2002 except for a little copy and paste work. I completely vibe coded three internal web admin sites for separate projects and used Amazon Cognito for authentication. I didn’t look at a line of code that AI generated any more than I would have looked at a line of code for a website I delegated to the web developer. I cared about functionality and UX.
I have not encountered an agent yet that I can trust in the same way.
Being shoes, offshoring, Webwidgets or AI generated code.
But we’re the shoemakers, not the consumers. It’s actually our job to preserve our own and our peers quality of life.
Cheapest good option possible doesn’t have to be the sweatshop - tho the shareholders of nike or zara would have you believe that - the labor movements of the 19th century proved that’s not the case.
How you define your proof is up to you. It might be a simple test, or an exhaustive suite of tests, or a formal proof. It doesn't matter. If the output of the code is correct by your definition, then it doesn't matter what the underlying code actually is.
If the spec is so complete that it covers everything, you might as well write the code.
The benefit of writing a spec and having the LLM code it, is that the LLM will fill in a lot of blanks. And it is this filling in of blanks that is non-deterministic.
Welcome to the usual offshoring experience.
To aim for a meeting of the minds... Would you help me out and unpack what you mean so there is less ambiguity? This might be minor terminological confusion. It is possible we have different takes, though -- that's what I'm trying to figure out.
There are at least two senses of 'correctness' that people sometimes mean: (a) correctness relative to a formal spec: this is expensive but doable*; (b) confidence that a spec matches human intent: IMO, usually a messy decision involving governance, organizational priorities, and resource constraints.
Sometimes people refer to software correctness problems in a very general sense, but I find it hard to parse those. I'm familiar with particular theoretical results such as Rice's theorem and the halting problem that pertain to arbitrary programs.
* With tools like {Lean, Dafny, Verus, Coq} and in projects like {CompCert, sel4}.
- I bootstrap AGENTS.md with my basic way of working and occasionally one or two project specific pieces
- I then write a DESIGN.md. How detailed or well specified it is varies from project to project: the other day I wrote a very complete DESIGN.md for a time tracking, invoice management and accounting system I wanted for my freelance biz. Because it was quite complete, the agent almost one-shot the whole thing
- I often also write a TECHNICAL-SPEC.md of some kind. Again how detailed varies.
- Finally I link to those two from the AGENTS. I also usually put in AGENTS that the agent should maintain the docs and keep them in sync with newer decisions I make along the way.
This system works well for me, but it's still very ad hoc and definitely doesn't follow any kind of formally defined spec standard. And I don't think it should, really? IMO, technically strict specs should be in your automated tests not your design docs.
I found it works very well in once-off scenarios, but the specs often drift from the implementation. Even if you let the model update the spec at the end, the next few work items will make parts of it obsolete.
Maybe that's exactly the goal that "codespeak" is trying to solve, but I'm skeptical this will work well without more formal specifications in the mix.
Yes and yes. I think it's an important direction in software engineering. It's something that people were trying to do a couple decades ago but agentic implementation of the spec makes it much more practical.
I have the same basic workflow as you outlined, then I feed the docs into blackbird, which generates a structured plan with task and sub tasks. Then you can have it execute tasks in dependency order, with options to pause for review after each task or an automated review when all child task for a given parents are complete.
It’s definitely still got some rough edges but it has been working pretty well for me.
https://codespeak.dev/blog/greenfield-project-tutorial-20260...
It is a formal "way" aka like using json or xml like tons of people are already doing.
You're telling me that I should be doing the agonizing parts in order for the LLM to do the routine part (transforming a description of a program into a formal description of a program.) Your list of things that "make no sense" are exactly the things that I want the LLMs to do. I want to be able to run the same spec again and see the LLM add a feature that I never expected (and wasn't in the last version run from the same spec) or modify tactics to accomplish user goals based on changes in technology or availability of new standards/vendors.
I want to see specs that move away from describing the specific functionality of programs altogether, and more into describing a usefulness or the convenience of a program that doesn't exist. I want to be able to feed the LLM requirements of what I want a program to be able to accomplish, and let the LLM research and implement the how. I only want to have to describe constraints i.e. it must enable me to be able to do A, B, and C, it must prevent X,Y, and Z; I want it to feel free to solve those constraints in the way it sees fit; and when I find myself unsatisfied with the output, I'll deliver it more constraints and ask it to regenerate.
Be careful what you wish for. This sounds great in theory but in practice it will probably mean a migration path for the users (UX changes, small details changed, cost dynamics and a large etc.)
The spec was essentially: "A user shares their professional background and a target job posting. The AI has a conversation to understand their experience, then generates a tailored resume and cover letter as downloadable PDFs. It remembers their profile within the session so they can generate documents for multiple roles."
That's it. No formal grammar, no structured templates for the conversation flow. The LLM handles the "how" — when to ask clarifying questions, how to reframe experience for different roles, what to emphasize based on the job description.
The interesting finding: natural language specs work remarkably well for applications where the output is also natural language (documents, advice, analysis). The formal spec approach makes more sense when the output needs to be deterministic code.
Try it if you're curious: https://super.myninja.ai/apps/6de082c7-a05f-4fc5-a7d3-ab56cc...
Is that really true? I haven’t tried to do my own inference since the first Llama models came out years ago, but I am pretty sure it was deterministic: if you fixed the seed and the input was the same, the output of the inference was always exactly the same.
1.) There is typically a temperature setting (even when not exposed, most major providers have stopped exposing it [esp in the TUIs]).
2.) Then, even with the temperature set to 0, it will be almost deterministic but you'll still observe small variations due to the limited precision of float numbers.
Edit: thanks for the corrections
No. Floating number arithmetic is deterministic. You don't get different answers for the same operations on the same machine just because of limited precision. There are reasons why it can be difficult to make sure that floating point operations agree across machines, but that is more of a (very annoying and difficult to make consistent) configuration thing than determinism.
(In general it is mildly frustrating to me to see software developers treat floating point as some sort of magic and ascribe all sorts of non-deterministic qualities to it. Yes floating point configuration for consistent results across machines can be absurdly annoying and nigh-impossible if you use transcendental functions and different binaries. No this does not mean if your program is giving different results for the same input on the same machine that this is a floating point issue).
In theory parallel execution combined with non-associativity can cause LLM inference to be non-deterministic. In practice that is not the case. LLM forward passes rarely use non-deterministic kernels (and these are usually explicitly marked as such e.g. in PyTorch).
You may be thinking of non-determinism caused by batching where different batch sizes can cause variations in output. This is not strictly speaking non-determinism from the perspective of the LLM, but is effectively non-determinism from the perspective of the end user, because generally the end user has no control over how a request is slotted into a batch.
Float addition is not associative, so the result of x1 + x2 + x3 + x4 depends on which order you add them in. This matters when the sum is parallelized, as the structure of the individual add operations will depend on how many cores are available at any given time.
And models I work with (claude,gemini etc) have the temperature parameter when you are using API.
It is absolutely workable, current inference engines are just lazy and dumb.
(I use a Zobrist hash to track and prune loops.)
So like when you give the same spec to 2 different programmers.
The entire thing about determinism is a red herring, because 1) it's not determinism but prompt instability, and 2) prompt instability doesn't matter because of the above. Intelligence (both human and machine) is not a formal domain, your inputs lack formal syntax, and that's fine. For some reason this basic concept creates endless confusion everywhere.
It’s not fine. I program using formal syntax precisely because I want the computer to do exactly what I tell it to.
What the Church–Rosser property/confluence is in term rewriting in lambda calculus is a possible lens.
To have a formally verified spec, one has to use some decidable fragment of FO.
If you try to replace code generation with rewriting things can get complicated fast.[2]
Rust uses affine types as an example and people try to add petri-nets[0] but in general petri-net reachability is Ackerman-complete [1]
It is just the trade off of using a context free like system like an LLM with natural language.
HoTT and how dependent types tend to break isomorphic ≃ equal Is another possible lens.
[0] https://arxiv.org/abs/2212.02754v3
Second, it's very easy to have some specs decided tractably, at least in many practical instances, but they are far too weak to specify most correctness properties programs need. You mentioned the Rust type system, and it cannot specify properties with interleaved quantifiers, which most interesting properties require.
And as for HoTT - or any of the many equivalent rich formalisms - checking their proofs is tractable, but not finding them. The intractability of verification of even very limited languages (again TQBF) holds regardless of how the verification is done.
I think it's best to take it step by step, and CodeSpeak's approach is pragmatic.
> First, it's not a question of decidability but of tractability
The question of decidability is a form of many-to-one, reduction. In fact RE-complete is defined by many-to-one reductions.
In a computational complexity sense, tractability is a far stronger notion. Basically an algorithm efficient if its time complexity at most PTIME for any size n input. A problem is "tractable" if there is an efficient algorithm that solves it.
You are correct, if you limit your expressiveness to PTIME, where because P == co-P, PEM/tight apartness/Omniscience principles hold.
But the problem is that Church–Rosser Property[0] (proofs ~= programs) and Brouwer–Heyting–Kolmogorov Interpretation[1] (Propositions at types) are NOT binary SAT, and you have concepts like mere propositions[3] that are very different than just BSAT.
But CodeSpeak doesn't have formal specifications, so this is irrelevant. Their example code output produced code with path traversal/resource exhaustion risks and correctness issues and is an example.
My personal opinion is that we will need to work within the limitations of the systems, and while it is trivial to come up with your own canary, I would recommend playing with [3] before the models directly target it.
Generating new code from a changed spec will be less difficult, specifically when the mess of real world specs comes into play. You can play with the example on CodeSpeak's front page, trying to close the various holes the software has with malformed/malicious input, giving the LLM the existing code base and you will see that "brown m&m"[3] problem arise quickly. At least for me if I prompt it to look at the changed natural language spec, generating new code it was more successful.
But for some models like the qwen3 coder next, the style resulted in far less happy path protections, which that model seems to have been trained on to deliver by default in some cases.
[0] https://calhoun.nps.edu/entities/publication/015f1bab-6642-4... [1] https://www.cs.cornell.edu/courses/cs6110/2017sp/lectures/le... [2] https://www.cambridge.org/core/journals/journal-of-functiona... [3] https://codemanship.wordpress.com/2025/10/03/llms-context-wi...
The computational limits imposed on program verification are independent of the logical theory used, and depend only on its expressive strength. Many if not most interesting program properties require interleaved properties (forall-exists or forall-exists-forall etc.) which are intractable to verify.
> But CodeSpeak doesn't have formal specifications, so this is irrelevant.
The lack of formalism also doesn't matter to the limitations on correctness. If you wish to know, with certainty, that a program satisfies some property, that knowledge has a cost. But both humans and LLMs write programs at least in part based on inductive rather than deductive reasoning, which cannot be rigorously given a level of confidence. That may not be what we want, but that's what computational complexity says we can have, so we're forced to work in this way.
We should complain about what we don't have, but demanding things we can't have isn't going to help. There's no fundamental reason why AI shouldn't, someday, be able to program as well as humans, but there are fundamental limitations to producing the software we wish we had. Humans can rigorously use deductive methods, with mechanical help, and AI could possibly use it, too. But there's no reason to believe AI could break the size barrier of such problems. People have been able to rigorously verify only very small programs, and maybe AI could do a little better if only because of its tenacity, but if we expect to produce perfect programs by any means, we'll be waiting for a long time.
I have no idea about codespeak - I was responding to the comments above, not about codespeak.
True.
> There's a vast chasm there we only need to even slightly start to bridge to get to far higher confidence levels than the typical human dev team achieves.
The word "slightly" is doing a lot of work here to the point of making it impossible to estimate. For example, the complexity classes P and NP are only slightly apart, and yet that's where a very practical barrier between feasibility and infeasibility lies. I don't doubt that one day AI may be able to write programs as well as humans, although nobody can estimate how soon that day will come, but nobody knows how wide the gap between that and "far higher confidence" is. Maybe there are fundamental computational complexity barriers in that gap that no amount of intelligence can cross, and maybe there aren't. Nobody knows yet.
What we do know is that anything humans do is possible - after all, we're doing it - and many things we need and humans can't do (including predicting nonlinear systems like the behavious of economy) no machine can do drastically better because of complexity limitations.
Slightly sarcastic but not sure this couldn't become a thing.
I use Kiro IDE (≠ Kiro CLI) primarily as a spec generator. In my experience, it's high-quality for creating and iterating on specs. Tools like Cursor are optimized for human-driven vibing -- they have great autocomplete, etc. Kiro, by contrast, is optimized around spec, which ironically has been the most effective approach I've found for driving agents.
I'd argue that Cursor, Antigravity, and similar tools are optimized for human steering, which explains their popularity, while Kiro is optimized for agent harnesses. That's also why it’s underused: it's quite opinionated, but very effective. Vibe-coding culture isn't sold on spec driven development (they think it's waterfall and summarily dismiss it -- even Yegge has this bias), so people tend to underrate it.
Kiro writes specs using structured formats like EARS and INCOSE (which is the spc format used in places like Boeing for engineering reqs). It performs automated reasoning to check for consistency, then generates a design document and task list from the spec -- similar to what Beads does. I usually spend a significant amount of time pressure-testing the spec before implementing (often hours to days), and it pays off. Writing a good, consistent spec is essentially the computer equivalent of "writing as a tool of thought" in practice.
Once the spec is tight, implementation tends to follow it closely. Kiro also generates property-based tests (PBTs) using Hypothesis in Python, inspired by Haskell's QuickCheck. These tests sweep the input domain and, when combined with traditional scenario-based unit tests, tend to produce code that adheres closely to the spec. I also add a small instruction "do red/green TDD" (I learned this from Simon Willison) and that one line alone improved the quality of all my tests. Kiro can technically implement the task list itself, but this is where agents come in. With the spec in hand, I use multiple headless CLI agents in tmux (e.g., Kiro CLI, Claude Code) for implementation. The results have been very good. With a solid Kiro spec and task list, agents usually implement everything end-to-end without stopping -- I haven’t found a need for Ralph loops. (agents sometimes tend to stop mid way on Claude plans, but I've never had that happen with Kiro, not sure why, maybe it's the checklist, which includes PBT tests as gates).
didn't have the strongest start, but the Kiro IDE is one of the best spec generators I've used, and it integrates extremely well with agent-driven workflows.
formal specification is no different from code: it will have bugs :)
There's no free lunch here: the informal-to-formal transition (be it words-to-code or words-to-formal-spec) comes through the non-deterministic models, period.
If we want to use the immense power of LLMs, we need to figure out a way to make this transition good enough
>* write text specifications
>* model transforms text into a formal specification
>* then the formal spec is translated into code which can be verified against the spec
This skill does just that: https://github.com/doubleuuser/rlm-workflow
Each stage produces its own output artifact (analysis, implementation plan, implementation summary, etc) and takes the previous phases' outputs as input. The artifact is locked after the stage is done, so there is no drift.
* This isn't a language, it's some tooling to map specs to code and re-generate
* Models aren't deterministic - every time you would try to re-apply you'd likely get different output (without feeding the current code into the re-apply and let it just recommend changes)
* Models are evolving rapidly, this months flavour of Codex/Sonnet/etc would very likely generate different code from last months
* Text specifications are always under-specified, lossy and tend to gloss over a huge amount of details that the code has to make concrete - this is fine in a small example, but in a larger code base?
* Every non-trivial codebase would be made up of of hundreds of specs that interact and influence each other - very hard (and context - heavy) to read all specs that impact functionality and keep it coherent
I do think there are opportunities in this space, but what I'd like to see is:
* write text specifications
* model transforms text into a *formal* specification
* then the formal spec is translated into code which can be verified against the spec
2 and three could be merged into one if there were practical/popular languages that also support verification, in the vain of ADA/Spark.
But you can also get there by generating tests from the formal specification that validate the implementation.