Joe Talks AI

About 16 months ago, I fine-tuned StarCoder 15B on 4× NVIDIA A10 24GB GPUs using QLoRA/PEFT. The results were genuinely good. But frontier models have improved so fast that I’m starting to wonder: Outside of highly specialized or regulated use cases, does the average AI team still need to fine-tune LLMs? Especially now that we have strong prompting, RAG, tools, and MCP. Who is still fine-tuning? and what are you getting that you couldn’t achieve another way?

@shorttimelines @thsottiaux @mylifcc Harness or model behavior, listen to yourself. Again do you even know how it all works? I mean from the data center back to your keyboard?

@JoeTalksAI @thsottiaux @mylifcc

找到一个降低5.6Sol token消费的办法： 在~/.codex/config.toml 中加入 model_context_window = 272000 model_auto_compact_token_limit = 240000 原理下面说：

The amount of hysteria, from builders "vibe coders" over the fact that the latest GPT 5.6 Sol model burns tokens too quickly is wild. They don't understand how it all works. Yes, I'm aware of the sub agent spawns with the same model and thinking. They're fixing that.

@shorttimelines @thsottiaux @mylifcc You have zero clue

@thsottiaux @mylifcc Tibo listen to the people man, it’s clearly burning more tokens than 5.5 at similar or even lower effort levels. This is either a harness or a model behaviour issue

@thsottiaux @mylifcc You have to feel for @thsottiaux it is literally the middle of the night and he's trying to make sure people don't set the wrong configs.

@mylifcc This is not correct. Do not do this if you do not understand exactly what you are doing. We do not charge extra above 270k context and the context threshold has been tuned for GPT 5.6 Sol to be perfect at the default limit.

@kidtsang Yeah I don't know how well it performs on actual intent or maybe it's just a solid model as a sub agent worker. I'll have to test it out

@JoeTalksAI I love that you're diving into GLM 5.2! It's impressive how AI can transform code quality. But I wonder how well it handles nuanced logic or edge cases. Refactoring is one thing, but can it truly grasp the app's intent?

Took GLM 5.2 for a spin, had it review a poorly vibe coded app example (around 45,000 lines of code) that I've built for testing new models and fully refactor it. Asked for it to review the codebase and make the recommended changes. Removed all the console.log with sensitive info, fixed JWT token leaks, fixed redundant code functions across multiple pages and a few other things. Took about 2 hours and cost around $5 in usage/tokens. I have to say it's probably 90% as frontier models, pretty impressive. #AI #GLM #Model #Coding #ZAI

@VsakoRA Thanks I appreciate that!

@JoeTalksAI What you are doing is wonderful; if you need a VPS server, I would be happy to offer a solution specifically tailored to your VIP/SIP project.

I've been experimenting with speech to speech models via WebRTC. Always enjoyed building voice apps since the early 2000's with VoIP, VXML, CCXML etc. GPT-realtime models were already pretty amazing. I'm excited to try the GPT-live models! There are so many areas where this is going to create a better voice experience, more realistic and less friction! #AI #OpenAI #WebRTC #Realtime

@kvickart People want outcomes, not a 30 min demo on another SaaS product. They don't want to click three levels deep, they want to explain what they need and spend less time actually using software to do that. The paradigm will shift again for sure.

I gotta be honest im starting to feel like software is baseline automated. Like sure there is obviously more that can be designed and invented and there are endlessly more things you can do with software. But the act of constructing it is practically approaching a completely non involved process. I was able to make a daw from scratch in an afternoon inside another piece of software today. Other people are making mmos. I feel like we're on the cusp of a fundamentally new paradigm in computing and to some extent even though i've embraced ai. I can't help but feel like im still thinking about it and approaching it in the completely wrong way. There's still a part of me that is clinging on to the old mentality of software is like some product you build and sell as some kind of service. Like realistically this time next year there is very unlikely to be the vibe coding process at all. I feel like you will just ask your computer to do what its capable of doing and it will just do whatever it can on demand. And there's a part of me that feels like this is the paradigm to mentally shift to, but i dont even really understand what that means i should be doing right now other than pivot to hardware and things computers just can't do

What does a vibe coder do when Anthropic goes down? Genuinely curious. They aren't going to go "look at the code" so..

@SeeLos Claude Cowork, Copilot Cowork, ChatGPT Work, see the trend?

I don't understand the point of the 'work' tab in chatgpt?

@skuffd @thsottiaux I wrote C# for 15 years, by hand. Built two projects from Thurs till now in C#, overall I don't see any major issues. C# backend though (API and Windows Service), React/Vite front end.

@thsottiaux .NET development Task runaway and over-reach Scope increase without approval

At different times you complained about speed, code slop, frontend quality, ... With each release we improve and GPT 5.6 Sol is ✅Fast and token efficient ✅Hardcore at back-end dev ✅Great at front-end ✅Does not use useEffect everywhere What is next?

My son wanted take GPT5.6 for a spin on creating an RTS 3D 1700's time period game. 4 hours later, the prototype was done. Music, Sound Effects, Game Mechanics, 20-30 min rounds vs PC. I guess I'm just shocked that you can build a simple prototype game that quickly with AI. #AI #Gaming #GPT56 #RTS

Especially a product landing page. Especially when you put this in your prompt: Use only: - Semantic HTML - CSS - Vanilla JavaScript - Original SVG authored during this run - Original CSS-generated graphics authored during this run I understand what you were trying to achieve but this also isn't representative of how most would prompt these models to get solid results. Your prompt is why the can designs look so bad.

Most frontier models still can’t make a frontend you’d actually ship. I gave GPT-5.6 Sol and the other top models one job: design a Diet Coke landing page. Same prompt. Same constraints. Fresh chat for every run. No shared references. No cleanup after. The screenshots below are the raw outputs. Which one wins? Guess before you see the model names.

What's your favorite harness today? Here's a fun fact, I've never used Claude directly through their product. Should I? #AI #Claude #Harness #GPT56

@uzairansar Terra not doing so hot eh? Classic middle child err model issues?

Some thoughts on the 5.6 models now that I've gotten to test them out: 1) Best Value: gpt-5.6-luna [xhigh/max] The base token cost is super low, so maxing out its effort level lets it brute reason it's way thru most tasks. 2) Best for Critical Code: gpt-5.6-sol [medium/high] When you absolutely cannot risk a broken build or a bad refactor. Sol [max] doesnt make too much sense - a tiny 2% performance bump isn't worth the cost increase. 3) Best Budget Baseline: gpt-5.6-sol [low] A reliable choice that can get most tasks done as long as they're sliced correctly and narrow in scope. Gives you great performance without eating up your budget. Maxing out a lighter model's effort level is actually a pretty cost-effective way to get great performance - although it might take more steps to get there.

It's been a busy week for the frontier model industry. OpenAI released their GPT-5.6 line of models (Sol, Terra, Luna) at the same time xAI released Grok 4.5 just after Anthropic was allowed to rerelease Fable 5, and now Meta has released their Muse Spark model. The advancement of these frontier models is significant. Beyond benchmarks, when you actually use them for the first time whether for coding or work related tasks you'll notice quite a difference. Grok 4.5 isn't Fable level but it's ability perform tasks at Opus level for a fraction of the token cost is very impressive. The GPT5.6 Sol model is pricey, not as pricey as Fable 5 but still up there. Both models have exceptional ability to reason, spawn sub agents, delegate work and accomplish what would have been a month to develop a fairly complex POC only takes hours. I say POC because there is a lot more that goes into scaling software for the real world. However, with harnesses being able to perform computer use tasks along with being coupled to these top end models, the results are speaking for themselves. Loop engineering and Harness engineering are growing in popularity for good reason. When you have the right context, hooks, memory system and properly laid out goals, the only limit is how many tokens you're ok spending to achieve results in days sometimes, perhaps weeks, but most likely not months. #AI #Software #OpenAI #HarnessEngineering

💀😭

@LoopOnChain That's what DCG hook is for. github.com/Dicklesworthst…

So apparently GPT 5.6 Sol might just randomly delete all of the files on your laptop 🚨 Here's how to safeguard yourself against it Paste this prompt into codex: Protect this machine from accidental permanent file deletion by Codex. Implement this as defense in depth, not merely as a written instruction. Requirements: 1. Inspect the current official Codex documentation for hooks, command rules, sandboxing, and approval settings before making changes. 2. Detect the operating system and identify its recoverable Trash or Recycle Bin command: - On macOS, use `/usr/bin/trash `. Do not add `--`, because macOS `/usr/bin/trash` treats it as a filename. - On Linux or Windows, verify an installed, reliable Trash or Recycle Bin mechanism before using it. - If no reliable recoverable mechanism exists, stop and explain what must be installed or configured. 3. Add a durable global Codex instruction: - Codex must never permanently delete a file or directory. - All removals must go to Trash or the Recycle Bin. - Permanent deletion requires the user to explicitly change this safety policy first. - Destructive Git operations such as `git clean`, `git reset --hard`, bulk checkout, or bulk restore must also be blocked. 4. Install a global `PreToolUse` hook that returns a hard `deny` before execution when it detects: - `rm`, `unlink`, `rmdir`, or `shred` - `find -delete` - `rsync --delete` - destructive Git cleanup or reset commands - common Python, Node, Ruby, or Perl deletion APIs - file deletion through `apply_patch` - delete-like MCP filesystem tools - nested destructive commands inside `bash -c`, `bash -lc`, `sh -c`, or `zsh -c` 5. The hook must direct Codex to the verified Trash or Recycle Bin command instead. Do not silently rewrite dangerous commands, because flags and shell expansion could be misinterpreted. 6. Add user-level Codex command rules that mark direct permanent-deletion commands as `forbidden`. Preserve existing rules and configuration. 7. Set safe global defaults when compatible with the existing configuration: - `sandbox_mode = "workspace-write"` - `approval_policy = "on-request"` Do not override managed policies or create conflicting permission configurations. 8. Preserve all existing configuration. Inspect files before editing, make narrowly scoped changes, and create recoverable backups of any configuration files that must be replaced. 9. Validate everything without risking real data: - Parse the hook and configuration files. - Feed simulated tool-call JSON into the hook. - Confirm dangerous examples are denied. - Confirm ordinary commands and the Trash command are allowed. - Test the Trash mechanism using a newly created disposable temporary file. - Run an end-to-end Codex test using `rm` with no arguments inside a read-only sandbox. Never test `rm` with a real target. - Confirm the command was blocked before shell execution. 10. Explain any limitations honestly. A Codex hook is a guardrail, not an absolute operating-system security boundary. 11. Check whether system backups are configured. Do not enable or modify backups without permission, but report clearly if no backup destination exists. 12. At completion, provide: - The files created or changed - The validation results - The exact one-time steps required to review and trust the hook - Any restart requirement - Any remaining risks Do not permanently delete anything while completing this task. If cleanup is necessary, move it to Trash or the Recycle Bin.

@theo The variability of harness's and LLM's. Harness hopping is a thing

gpt-5.6-sol is meaningfully better in Claude Code than in Codex I'm going to crash out so badly over this

@kr0der Medium

it's been 1 day now - what reasoning level is everyone using for GPT 5.6 Sol?