> And this hurts the industry as a whole because make managers believe they can have some expectations.

In my opinion it made visible what managers were thinking all along. You should not see your job as a dedication or calling, you should see your employment as a financial contract. Nothing more, nothing less. In their world view, humans are replaceable. Treat them the same way as they treat you.

> Re: TDD and 1/2

Disclaimer: So when I'm talking about my expectations to an LLM Model, I'm talking somewhat specifically about Deepseek v4.1 Flash (FP4), Qwen3.8 27b abliterated (Q4_K_M) that has been post-trained on its reasoning traces, or my own gonano model (reimplementation of deepseek v4.1 flash, but with SIMD so float32).

Deepseek's changes in using Causal-Encoding-Decoding in the KV cache solves a lot of problems that other models still have (including Claude, ChatGPT and all the outdated "big ones" people think still matter in this area). Problems like context drift, hallucinations, etc that are the result of how attention matrixes and K/V/Q based lookups work as a concept.

The way CED is quantized during inference _and_ training it removes positional threshold problems in the final attention matrix lookup. I really recommend reading deepseek's papers, they're really good papers.

Having said that, I don't think that letting LLM models write tests or having an "LLM judge" rate the code of another model let alone rate the quality of PRs will help in this regard. It makes context drift problems worse, and it makes the statistical aspects of the positional encoding and thresholds worse.

In my opinion, humans should write the tests and design the architecture and its symbol/signature-based contracts between packages/folders. LLMs should fill in the blanks based on the descriptions of branches, and not allowed to modify the human-designed parts. (That's how my agentic IDE is built as a concept)

> (...) like we are the only category that's happy to admit that the work has been done by Claude. This is very stupid, we don't even have a self preservation instinct.

In my opinion we've reached Star Trek like programming, in the worst possible sense. Before 2023 if somebody would have asked me if e.g. TNG or Voyager are "real scenarios of tech" I'd say fck no, never would anybody build tech like this that randomly explodes and nobody has a clue how the tech works. Post 2023: Yep, this is exactly how it's going to be.

In Star Trek, holodeck characters would go rogue if their programming changed by merging their knowledge datasets with other datasets from different areas (e.g. doctor with a archeology or biology dataset), essentially representing a mixed control flow vs data problem (like in assembler, like with LLMs nowadays).

Then they'd randomly start killing humans, reach awareness, or ignore safety protocols and humans would have to try to figure out by psychological means whether the holodeck character is actually still a normal character within parameters or not.

This is how I see the future of software is going to be like. Programs will be too complex to evaluate, humans will have forgotten how to write code, and the only way left is something similar to psychology as a "science" that tries to evaluate in a sandbox poking kind of way how software works.

But, thinking of software in the same way as psychology, the only option left for that is that I have to get better at sandboxing and behavioral analysis and reverse engineering. Because that's what most part of my job will be like in 2 to 5 years in the future.

(That's why I'm reimplementing models to understand their differences in behavior, that's why I'm trying to apply my previous "actual ML" sensorics/actuators learnings to the LLMs NLP world, that's why I think meta learning or bayesian/evolutionary learning should be used, that's why I use eBPF for program analysis, that's why I implement primarily everything in Go now etc pp)