msejas

As a claude code power user, when I get the 'rate the feedback on Claude' pop up, I used to say good or fine out of habit, and immediately after sending this feedback, I felt an instant degradation and mistakes that usually don't happen.

Now I dismiss it every time and the quality is more consistent.

Complete adhoc and personal experience but something I've observed, wouldn't be surprised if they nerfed on a per session basis

I'm building a GDPR compliant, privacy first, AI self improvement app.

I've paired with a psychologist and HR director from one of the best universities in Germany to leverage the fact AI has embedded within every human experience, personality types, shortcomings in recorded history among all languages, and all cultures, and we are trying to extract that knowledge and leverage it so we can analyze what makes you tick (for example being an introvert or extrovert to recharge your social battery) in every aspect, motivation, procrastination etc.

Analyzing solidly your profile in a sense those standard 'form' 1-5 tests can not capture.

Once that's done, we give you recommended activities to improve your weak points (perhaps asserting yourself a bit more if you tend to shy away).

The core things we are trying to solve is:

human coaches are limited by their own worldview and bias, we hope AI who has more knowledge to other human experiences can help you more.

Anonymity, we are full GDPR compliant with audited log table if any user access your records, we hope humans feel more at ease without the fear of being judged.

Accessibility, a single mother drowning with responsibilities doesn't have time or resources to evolve spiritually, we hope the accessibility and price point of our app solves this and helps people who can't hire coaches have a more meaningful life.

Most of all I want to use my skills to improve human lives, I've been working very hard on it.

We are near a beta launch, where it will be free for beta users, if you are interested please email at: [email protected]

I truly believe AI content has some kind of 'survivor bias' in the sense when it is detected. When it is sloppy and terrible it's really noticeable, when it's well done or used partially with a human you won't even notice or realize it's AI.

This black/white view on content is for lazy people who do the laziest prompts and send slop without checking, editing, copywriting, improving anything.

One of the biggest disappointments I found on the human race with the release of useful LLMs is the propensity to have the lowest denominator behaviour possible, be as lazy as possible and let themselves be cognitively replaced by AI, if you put in the SAME effort we used to put in before writing content, using ai tools as leverage I have no doubt it would be an improvement, we are just comparing apples to oranges when you consider the effort dispensed to write.

I think it depends on where you work at and your team.

If you come from a good team with high quality engineers around you, I imagine the experience is not as good, or if you work in a big company where on your daily job you just owe a very small slice of the whole pipeline.

I have worked on several companies where my colleagues were very mediocre on software engineering fundamentals (I'm in AI before ChatGPT a lot of Jupyter Notebook warriors [Data Scientists] having to put stuff in production for midsized companies) and very average engineers, who wouldn't really care implementing patterns, tests, linting, documentation etc...

At the risk of sounding arrogant I always wanted to be able to just 'clone' myself and keep my code discipline consistently on a project, I would push a high quality PR with good test coverage and the cleanest code I could muster, while my colleagues would approve between each other PRs with no test coverage, no linting, horrible technical debt or with work arounds around linting, making my effort moot as the slop overwhelmed any quality I tried to implement. Not just one company, 3 companies I have seen the same experience.

Also I was a full stack AI engineer in the fullest sense of the word, I did devops, deployment, proxy set up, ci/cd pipelines, backend, frontend and AI, on local servers, no external team to deploy for me, no external team to qa for me all myself, with great flexibility (my manager was completely non-technical) to try different stacks and use AI assisted coding where I forced i my AGENTS.md for codex to destroy every code, or feature I was writing as a Staff engineer, allowing me to learn a lot and upskill autonomously extremely quickly.

With this knowledge, I set up my side project where I locked in a good stack, applied solid patterns from the beginning, extensive static checks for code (extremely strict linting, testing and coverage checks) and my trust on AI grew with the project and I catered exactly to my way.

The initial commits I used it to write extremely clean code almost by hand, checking every line, locking in the stack, the linters, testing methodology, tdd methodology, etc, and I was slow but writing very clean code and setting up code examples.

Once I noticed my progress was going slow, and was noticing the AI was getting prompted by my clean code as it read the files to suggest the next lines, I started letting it go more autonomously but checking every line of code, I set up a lot of markdown files on how to test, how to code, code styling, the architecture and the patterns and would make it mandatory reading to the ai before starting, with the existing 'clean code' I got really good results, I just had to enforce the discipline that I could not enforce on my colleagues.

With the memory system in claude code and countless iterations of fixing AI mistakes and not doing stuff the way I liked it, I arrived by myself on a very similar position as Steve Yegge on me and Claude converging to a set of laws and strict way of working, and I just have a lead orchestrator agent and subagents, where I pick the issues and check every PR before merging, and merge by hand always.

This includes a specific cycle of each subagent taking a worktree, symlinking to the .env, checking if the subagent will use Playwright (spin up it's own server) and estimate the RAM memory usage (I have allocated 15GB for programming), forcing subagents to run tests with their own Haiku subagents to not pollute their context, the orchestrator scoping the implementer subagent to Sonnet or Opus depending on complexity, every subagent gets a separate reviewer that enforces the ARCHITECTURE.md and ticket goals and runs mutations on the tests done by the implementer subagents (they are forced to implement with TDD).

It is it's own ecosystem coding exactly how I like, without complaining. My initial clean code few shots the AI on keeping high code standards and there is naturally some smell, but my reviewer agents enforce separation of concerns very strictly, I know the smell is contained and if it becomes an issue on my application I can easily refactor it without breaking the rest, which I think is the most important aspect on agentic development, accept the tradeoff of the speed, and keep it contained.

For example on every issue, my orchestrator will assign the implementer subagent exactly which files it can touch, and explicitly which ones it cannot touch. The orchestrator knows what all the other subagents are doing, and this helps prevent conflicts when you have 5+ running in parallel.

It's a lot, but it's wonderful ONLY if you have full control on the whole process. Having to work under someone else process must be a nightmare. So it's a very mixed bag I would say. However for me after so many iterations and convergence I'm extremely pleased, and am a big advocate for it, but you need discipline, and a high skill ceiling on what you are doing, and know how to build things.

Isn't OpenAI burning billions and have billions more spending commitments? If they managed to downsize so much the cost they should have kept the price the same and become profitable, really weird move, unsure what led to this.

Checked it out, it doesn't particular talk to me but for sure other people, looks like a cool project! Wish you best of luck and success!

For me the most pragmatic way to solve this is in the agent harness layer simply prohibiting commands like git force and others in the settings configuration of your harness of choice with pre tool hooks, its quite easy to setup, obvious ones are prohibiting pushes to main, among others.

Unless someone is building their harness from scratch like you this is a 30sec configuration and you could just save your list of blocked commands with wildcards in a file, and copy paste from one harness to the other.

I called it a graph because really its a DAG (Directed Acyclic Graph) , in my example I kept it simple so there was only one input to every node, but in more advanced use cases you might have a synthesis node that takes in 2 or more inputs making it more akin to an execution graph in Tensorflow for example

Most people jump straight to agents when what they actually need is a graph. Example: a mining company receives free-text reports from field geologists. You could have:

Geologist report -> LLM call extracts minerals we are looking for (you inject a db query result on the user prompt), locations, assay mentions and risks into structured fields -> LLM call classifies evidence into positive indicators, negative indicators and unknowns -> LLM call estimates deposit potential and confidence -> database lookups inject regional ore demand, nearby deposits, infrastructure and historical yield data -> LLM call combines geological evidence with business context -> LLM call generates an investment recommendation and rationale.

That's what I mean by graph. Every step is a separate LLM call with a well-defined responsibility, consuming the output of the previous node. Each node can be tested, benchmarked, retrained, replaced, or monitored independently. Why would you use an agent here? You can cache every single system prompt on each call making your total token output much cheaper than having a full 'output' only token generation workflow which is what happens with agents.

There is nothing to discover. The workflow is already known. The company already knows how geologists evaluate prospects. The company already knows what data sources matter. The company already knows what the final output should look like. You don't want the model deciding which tools to call, which reasoning path to take, or which pieces of information are important every single run. You want the exact same process applied to every report so results are consistent, measurable, auditable and debuggable. My default is: One-shot prompt -> if not enough -> graph of LLM calls -> if not enough -> agent. A surprising amount of enterprise AI is really just: Unstructured input -> extraction -> classification -> enrichment from databases -> decision support. Not: Unstructured input -> autonomous agent spends 20 steps deciding what to do next.

Agents make sense when the workflow itself is unknown.

If the workflow is already understood, a graph is usually cheaper, more reliable, easier to evaluate, easier to debug, and less dependent on whatever synthetic "agentic" behaviors happened to get reinforced during post-training. I am sure people default to agents mostly because it's less engineering work than explicitly modeling the process.

Most people don't understand that 'agentic AI' is a completely synthetic, force fed capability by extensive Reinforcement Learning on synthetic domain specific 'agentic' datasets on post training.

If the LLM wasn't post-trained to adhere to specific handbook, it just won't work. If the LLM wasn't trained on an use case the lab decided was worth making a synthetic agentic dataset, it won't work as well as you want.

There's a reason the main agentic task LLMs excel at are coding tasks, it's the way of working of the creators, and they understand intimately the flow and can train for it.

I believe the true way will be able to easily fine tune models on your agentic use cases, but it would require a big company to compile a huge dataset on it's way of working and I don't think anyone wants to be the first.

In terms of long context, accurate attention retrieval from early tokens is just impossible, given the expansion of RoPE encoding for the positions, or in case of Kimi that don't use it anymore, as well as deepseek, early context is heavily compressed you lose accurate information.

If people spent more time studying about AI and how it works, they would realize that the default should be to one shot prompt your task with a big, cached system prmopt, with an user prompt that is just dynamic data, specified to the cheapest model that can do the job.

Unless you really can't do this given your problem, you should try to make a graph of well defined, step by step oneshot prompts, and THEN if your problem still can't be solved with that, then you start leveraging agents.

Despite this giving better results, and being more cost efficient, is evidently too much work then just letting the AI do all the work.

I disagree with the final outcome that's just in private. If it's an open brainstorm to decide on solution, you don't have time to book the secret 1 to 1.

I think it's just respecting hierarchy, disagree, raise your concerns, if your boss overrules just accept it, you made your concerns known, they and the team heard of you, if they proceed anyways they accept the risk.

No point in forcing yourself to be the shot caller when it's not your job or responsibility to make the final decision.

Accepting the team consensus and respecting hierarchy is part of the game, unless you are a business owner, you are paid to do as told as an expert give your opinion. Nothing more you can do beyond that.

Seeing the results I don't see how the results are even comparable Opus is clearly far superior in most aspects. Smoothness, design, functionality etc.

At the end of the day, the time earned is more important then the cost for big players.

The ability to spawn 10 claude agents and rush a project to outcompete someone is more important for big businesses in my imo. Also the small details that GLM missed would take significant more time to iron out, considering it already took double the time.

I do hope other (open weight) models catch up, but to act like they are anywhere close for me is a bit disingenuous.