Hacker News: computerex

New comment by computerex in "GLM-5.3: Frontier coding with emergent cyber capabilities"

computerex — Sat, 15 Aug 2026 04:30:37 +0000

The LLM is the core model, but the harness has the prompts/tool definitions, guidance/recovery/correction code. The harness itself is the agent, because same model may perform vastly differently on different harnesses. Agent is the system working as a whole, harness+llm.

New comment by computerex in "GLM-5.3: Frontier coding with emergent cyber capabilities"

computerex — Fri, 14 Aug 2026 19:41:54 +0000

The harness is the agent. LLM's can be asked to output things in JSON for example. The LLM then literally asks for things like "execute this cmd" or search/replace this string. The LLM outputs text, but in a deterministic format that can be parsed. The harness calls the LLM, exposes tools, executes tools the LLM asks for, gates tool use based on security controls. It's the runtime that the agent uses to do work.

New comment by computerex in "DeepSeek V4 Pro 0813"

computerex — Thu, 13 Aug 2026 18:25:04 +0000

Yes, and the reason why pass@k exists is because of self-consistency. There is no guarantee for right answer to be selected or for the LLM to correct itself. While I agree pass@1 is a useful metric, I'd be more interested to know pass@5 so I can better compare the results.

New comment by computerex in "DeepSeek V4 Pro 0813"

computerex — Wed, 12 Aug 2026 20:59:06 +0000

An agent doing a task even with multiple back to back calls like normal without an example is zero shot. An agent doing a task with 1 example is one shot. An agent doing a task with a few examples is few shot. I don't think you are correctly using these terms.

The multiple back to back LLM calls are done on accumulating context, so if there is a sampling error it could throw the entire session out of whack, because LLM's build on the previous context.

It's actually meaningless to argue, one could simply sample more than 1 times and let the numbers speak for themselves.

New comment by computerex in "DeepSeek V4 Pro 0813"

computerex — Wed, 12 Aug 2026 19:25:41 +0000

They are causal autoregressive models, the output is sensitive even to the implementation nuances in inference. Even 1 token that's badly selected could throw off the entire answer.

New comment by computerex in "DeepSeek V4 Pro 0813"

computerex — Wed, 12 Aug 2026 17:39:53 +0000

Repeat the test like 5 times for each model and see the results.

New comment by computerex in "Woman Pulled over at Gunpoint Twice After Flock Camera Glitch"

computerex — Tue, 11 Aug 2026 18:24:22 +0000

While I agree with need more and better gun control, American police officers are abhorrent for a very simple reason: lack of basic training, discipline, and practically zero screening/qualification process.

Anyone can become a cop in 6 months. Not everyone is temperamentally setup to be a great cop. That, there, lies the rub.

New comment by computerex in "Advancing the price-performance frontier with GPT‑5.6"

computerex — Thu, 30 Jul 2026 19:15:06 +0000

Absolutely. Although DeepSeek started announcing "Peak valley" pricing which started making me nervous. I have spent $50 usd in July on deepseek and for that much spend I got SO MUCH mileage.

I feel perfectly content in using pay as you go pricing with deepseek. On the other hand, although Anthropic's models used to be my bread and butter for personal work, they are simply too expensive to reach for these days.

New comment by computerex in "Codex Security"

computerex — Wed, 29 Jul 2026 01:19:29 +0000

Not really. I test the output thoroughly, I examine the thinking process, I go through the diff to see if anything jumps out but I my thinking process/the way I work has changed. Low level programming thinking has gotten atrophied it seems.

New comment by computerex in "Codex Security"

computerex — Wed, 29 Jul 2026 01:16:44 +0000

I think it's because python is far more approachable/ubiquitous than go/rust. It's the entry level language for many people from all disciplines of life. Scientific community uses it, data science uses it.

Golang/rust however are very convenient to distribute. Small, portable, fast exe's are very nice. With agentic coding golang/rust are now accessible to a lot more people.

New comment by computerex in "Is this the end of the once-mighty GoPro?"

computerex — Sat, 18 Jul 2026 16:04:14 +0000

People know it’s happening. What do you expect an average consumer to do about it? Pay more out of pocket due to the potential national security risks?

New comment by computerex in "Kimi K3: Open Frontier Intelligence"

computerex — Thu, 16 Jul 2026 17:06:23 +0000

I know GLM is relatively expensive and so is Kimi, in comparison to those DeepSeek V4 pro and flash are a godsend and are absolutely good value.

New comment by computerex in "Google Books (or similar) all book scans – $200k bounty (2025)"

computerex — Sat, 04 Jul 2026 19:04:24 +0000

That’s not remotely true. They did distillation as a cheap solution to the cold start problem. You need data/trajectories to hill climb to higher capabilities. All large Chinese labs do RLAIF.

New comment by computerex in "ZCode – Harness for GLM-5.2"

computerex — Thu, 02 Jul 2026 02:48:44 +0000

Why not? They are relatively easy to make so why not. Even I made one: https://github.com/computerex/z

New comment by computerex in "ZCode – Harness for GLM-5.2"

computerex — Thu, 02 Jul 2026 02:47:02 +0000

Not really. I have spent 163M deepseek v4 flash tokens in July and it literally just started.

https://i.postimg.cc/MHhgwsv0/image.png

New comment by computerex in "Leanstral 1.5"

computerex — Wed, 01 Jul 2026 00:05:06 +0000

It's for mathematics. There is this programming language: https://lean-lang.org/

If you can express a solution in Lean you can formally prove or disprove it. Formal verification is making a debut in traditional engineering toolkits.

New comment by computerex in "Claude Code is steganographically marking requests"

computerex — Tue, 30 Jun 2026 22:48:13 +0000

You may be ok with the harness doing 100 things that are not what I am using it for. But none everyone is, and it’s hardly hysterical. Perhaps you are simply careless.

New comment by computerex in "Claude Code is steganographically marking requests"

computerex — Tue, 30 Jun 2026 22:46:52 +0000

I made my own! https://github.com/computerex/z

New comment by computerex in "Nano Banana 2 Lite"

computerex — Tue, 30 Jun 2026 21:47:23 +0000

Llama was great for research and propelling the opensource community but they haven't been a serious competitor in a while. Whereas Google is one of the top 3 AI labs in America.

New comment by computerex in "Claude Code is steganographically marking requests"

computerex — Tue, 30 Jun 2026 21:18:02 +0000

I don't want my harness doing sneaky stuff like this. I don't want my harness data mining me. I want my harness to implement the agentic loop and I want it to be transparent.