Sabitlenmiş Tweet
I asked @coderabbitai 's VP of AI, David Loker: "How to build an agent that actually works."
What he said is confirmed by research.
infoworld.com/article/414152…
English
Andrew C. Oliver
27.3K posts
Andrew C. Oliver
@acoliver
Startup Marketing Exec Professional Vibecoder Infoworld column https://t.co/gf3nADYrVe. Opinions = Mine.
The World Katılım Haziran 2008
1.6K Takip Edilen1.9K Takipçiler
The great irony is that if your RAG returns straight, unsorted, irrelevant garbage, the model will likely disregard it.
English
Reminder: arxiv.org/html/2505.1876…
The better your RAG thing is, and the more plausible the "semantically" (but not actually) relevant result is -- the more it will screw everything up.
Andrej Karpathy@karpathy
One common issue with personalization in all LLMs is how distracting memory seems to be for the models. A single question from 2 months ago about some topic can keep coming up as some kind of a deep interest of mine with undue mentions in perpetuity. Some kind of trying too hard.
English
This is a scam. However, it did make me double-check by having llxprt-code inspect it. Over SMS was pretty unlikely. Don't go to the (russian) link but pretty good job wasn't it?
English
So @OpenAIDevs what is it with 5.4. Same harness same workflow and 5.3-codex is just more agentic and follows instructions better and asks for fewer needless confirmations. 5.4 It's like clearly smarter, but it needs some kind of fine-tuning. Is a 5.4-codex coming soon?
English
No one can be Silicon Valley without letting the IC get started there and letting them create a number of superfund sites then pass the money back and forth for decades. Its not the fairs or the rest its a monied investor class that makes everyone move there for funding.
dax@thdxr
English
Support increased mass surveillance! For the children!
tomshardware.com/software/opera…
English
If your secret sauce is a really special system prompt, you have bigger problems than the model leaking it. #nomoremoats
Dan McAteer@daniel_mac8
No prompt is safe. This is a real problem if your prompts are highly optimized and you invested a lot of effort into them. What can you do?
English
@svpino @Yuchenj_UW I just added this to my settings.json back when I used claude code:
"includeCoAuthoredBy": false
LLxprt Code doesn't do this because it's obnoxious.
English
@Yuchenj_UW I really hate that Claude does this. I had to write my own skill + hook to prevent it from doing this.
English
I noticed something interesting:
Claude Code auto-adds itself as a co-author on every git commit. Codex doesn’t.
That’s why you see Claude everywhere on GitHub, but not Codex.
I wonder why OpenAI is not doing that. Feels like an obvious branding strategy OpenAI is skipping.
English
Maybe it can but they canceled its contract.... 😏
Skindie@skindie___
If Claude code is so good why can’t it open the Strait of Hormuz
English
@trq212 @Amank1412 So k2.5 is really not bad. Use it to drive and use gpt-5.3-codex or 5.4 as a reviewer and hard problem solver. Anthropic is too behaviorally unreliable as a company. Not about their models but they cut everyone off, then up the quota, down, and consider use abuse.
English
@Amank1412 hi apologies, send me your email?
we're getting massive amounts of fraudulent requests so there are some false positives but we're working on bringing that down as much as possible
English
So GPT-5.4. I think it was rushed out the door a bit. It is clearly smarter but also clearly takes more effort to align. GPT-5.3-codex isn't as smart but will follow directions and is more agentic automatically. It feels like it was an aborted get it out the door finetuning issue
English
Discovered an interesting security issue... Keyring use in a node-based application has "node" ask for permission. So effectively, all node apps have access. The same thing in a Rust application is only for that signed version. Seems like this would be solved no?
English
Andrew C. Oliver retweetledi
was messing with the OpenAI base URL in Cursor and caught this
accounts/anysphere/models/kimi-k2p5-rl-0317-s515-fast
so composer 2 is just Kimi K2.5 with RL
at least rename the model ID
Cursor@cursor_ai
Composer 2 is now available in Cursor.
English
Hey @NarcosNetflix - for next series, do it in Southeast Asia but with GPUs instead of cocaine?
English
Good idea💡+implementation 💪 = 🥇
Good idea💡+implementation 💪 +open source = 🔥
Good idea💡+implementation 💪 + blockchain = 💩
Sorry that I'm not sorry...
English
Btw it repeats that same atructure many times: statement. Not opposite. Statement but shorter.
I didnt check the accuracy because I didn't read it because it was written this way.
English
It is fine to use LLMs in your writing. It is not fine to use them with no style guide without fixing their default writing style. If you didn't fix that did you check the rest? Also it makes people's eyes roll. Not read. Not look with interest. Roll.
towardsdatascience.com/the-multi-agen…
English