No, unreviewed was the previous policy. Now it’s none at all.
6,586 karma · joined March 13, 2013
Then before submitting if I ask an LLM to review my code and it proposes a few changed lines that are more efficient. Should I then
- Leave my less efficient code unchanged?
- Try to rewrite what was suggested in a way that’s not too similar to what the LLM suggested?
It also doesn’t necessarily mean he was a yes man. Often in these situations people spell out their confidence levels plainly and directly to such execs and it just bounces off.
I’m also seeing below people suggesting he could have publicly voiced his concerns, but that probably wasn’t even a legal option for multiple reasons.
Maybe, but for lots of scenarios even M5 could still benefit from being an order of magnitude faster.
AI, dev, some content scenarios, etc…
Hope they’re addressing that at the same time.
No, I wouldn’t think.
The post implies flaws in the original paper, then at the very end seems to concede it was technically fine just should’ve been up front about entanglement.
The post should be edited to be more upfront about what he realized after writing it.
No published results, missing details/lack of transparency, quality of the research is unknown.
Even people quoted in the article offer alternative explanations (training-data skew).
Commented after I saw this added in today’s release notes: “initial MCP interface and docs”
Updates (v0.36) https://github.com/openai/codex/releases
Typically less than 1% of training data is memorized.
“each cell birth plays a harmonic note and each death plays a complementary tone”
How are you deciding which notes to play?
Is it a function that somehow depends on generations or position?
That would result in a brittle solution and/or cat and mouse game.
The text that goes into a prompt is vast when you consider common web and document searches are.
It’s going to be a long road to good security requiring multiple levels of defense and ongoing solutions.
This will require improvements to vision models, RL frameworks, etc, but will be interesting to see how much it can broaden current abilities.
It’s easy to become jaded with so many huge models being released, but the reality is they are still from a relatively small group of countries.
For example India has no indigenous models this big despite having a world class talent pool.
Is there an example you can show that tended to fail?
I’m curious how token constraint could have strayed so far from your desired format.
RAG is Dead, Context Engineering is King
It’s difficult to read posts that rely so heavily on AI generated prose.
Everything’s a numbered/bulleted list and the same old turns of speech describe any scenario.
That aside, what’s really keeping this from being useful is showing some results. How well does this approach work? Who knows. If the data is sensitive, seeing it work on an open source repo would still illuminate.
Also, we hear lots elsewhere about the limitations of relying on embeddings for coding tools, it would be interesting to know how those limitations are overcome here.
There are a few ways to approach the problem. Pre-training on longer context lengths I’ve already mentioned. Fine-tuning techniques (like LongRoPE) I’ve already mentioned.
Inference time context extension tricks I didn’t mention because the papers I’ve seen seem to suggest there’s often problems with quality or unfavorable tradeoffs.
There’s no magic way around these limits, it’s a real engineering problem.
The largest context window a model can offer at a given quality level depends on the context size the model was pretrained with as well as specific fine tuning techniques.
It’s not simply a matter of considering increased costs.
- If Bitrig can do things like let your app have editable data, or make rest api calls, etc. it would be nice to see a little more in the demo video.
- Sounds like you don’t have an expensive mic/studio for recording, but luckily this is free and can work wonders for your audio: https://podcast.adobe.com/en/enhance
Best of luck with the launch.
Yes, if you want code to be maintainable over time, or to be able to evolve as a coherent foundation that lends itself to a long term direction and vision.
However it’s also the direction given to the AI enhancement - it’s steered toward buzzword/hype style.
“I managed a swarm of AI agents…”? How about instead just “I ran multiple instances of Claude Code and the results seem promising”? No swarms necessary.
Tim is the only side willing to publish papers and let them be peer reviewed.
He’s also the only one willing to engage on the merits of the debate. Eric has/will not.
If it doesn’t work for you on l33t code problems, what techniques are you finding more effective in that case?