Oday Brahem

build for the love of building

[Andrej Karpathy @karpathy]

One pattern I find useful for working with LLMs is a nice long ramble session. Sometimes the LLM needs more bits to understand what you're trying to achieve, but you're too lazy to type them. In these cases I like to lean back, switch to /voice and just ramble for like 10 minutes, total mess, anything goes, full stream of consciousness. Sometimes I declare it up top, something like "switching to speech recognition sorry for any typos...". Sometimes I turn it into a small interview of a few turns. But I find that the LLMs are somehow very good at reconstructing long incoherent rambles and often their echo of your own tangle of thoughts comes out quite a bit cleaner than what you started with. The result is that you improve the mind meld and have to correct things less from that point on.

[Oday Brahem @odaybrahem]

@karpathy This is very true and I find over time the rambles get better. I think we’re so accustomed to proper speech with other people that we forget that LLMs don’t need perfect speech. This is also why I find it funny when people spend so much time correcting the typos in their prompt.

[Oday Brahem @odaybrahem]

@Brusco_RL @anduriltech Congrats Nicolas! Super impressive. I’m rooting for you.

[Oday Brahem @odaybrahem]

@supabase An open source agent for mac that can see your screen and help you in context with memory.
https://t.co/wuD3aXc7xe

[Oday Brahem @odaybrahem]

@Brusco_RL I did it by using DR and training on a rented H200 with 8K parallelization.

[Oday Brahem @odaybrahem]

@Brusco_RL I did in the past transferring from JAX environment to Gazebo to Microsoft air sim but had more telemetry than given in VQ2.

[Nicolas Brusco @Brusco_RL]

Since Andruil has't released the leaderboard yet, has anyone else completed VQ2? I'd like to compare notes

[Oday Brahem @odaybrahem]

I love RL. One of the few areas where even frontier models will still struggle. Keeps it interesting.

[Oday Brahem @odaybrahem]

@thsottiaux More persona and life context. Proactive notifications would be nice too. Would love it if built in it could give me updates on recent work being done on other machines. Suggest new actions.

[Oday Brahem @odaybrahem]

@plauxai Yeah might be worth only releasing it that way for now. Source only no signed app. Good call out. First mac app release. 😅

[Oday Brahem @odaybrahem]

Anouncing Sonitura reinvented. This one is open source and costs nothing if you have a ChatGPT subscription. Leveraging your existing subscription this tool allows you to click (fn + ctrl) and ask an AI agent to help you with whatever you’re looking at on your screen. Keeps context, has long term memory and is something I’ve been using for the last couple of days with huge success. It was fun to build too! Since it requires a lot of permissions from Mac to work well I think this kind of tool needs to be open source.

It can even type for you when talking to your agents, if you’re ever talking to a coding agent and need help responding, Sonitura will type a response directly in context.

I’m on an open sourcing roll right now. Excited to share all the other agent tooling I’ve built over the last year with my 6000+ commits.

https://t.co/sMiSFr9obc

[Oday Brahem @odaybrahem]

@thsottiaux full repo https://t.co/OyD2fRPuZ9

[Oday Brahem @odaybrahem]

@thsottiaux Also built https://t.co/dh5lyW4e1G to meter your AI agent spend and put caps. Open source. Detects how much subscriptions would cost if charged at API rates. It’s been a busy week of building.

[Oday Brahem @odaybrahem]

@travisk @DavidSacks @elonmusk For people saying this is not real, I was able to reproduce it as well. If they haven't patched it you might be able to do the same.
https://t.co/KUPR20rkOZ

[Oday Brahem @odaybrahem]

Okay I saw someone posting this showing that kimi is claiming its claude so I tried it myself, wild!

[Oday Brahem @odaybrahem]

@lior_eth @travisk @DavidSacks @elonmusk I reproduced this exact response. It said it was Claude.
https://t.co/KUPR20rkOZ

[Oday Brahem @odaybrahem]

Okay I saw someone posting this showing that kimi is claiming its claude so I tried it myself, wild!

[Oday Brahem @odaybrahem]

@thsottiaux It also keeps context with your previous agent logs on your machine so it has some familiarity with what you are working on.

[Oday Brahem @odaybrahem]

@thsottiaux Built this A voice assistant that sees your screen, hold a key, ask, and it answers out loud on the ChatGPT plan you already pay for. Local speech-to-text + neural TTS, on-screen agent, private local memory. Open source, macOS.

https://t.co/wuD3aXc7xe

[Oday Brahem @odaybrahem]

@PimDeWitte I second this. I even tried it myself and had the same experience.

https://t.co/MaMcp2WktD

[Oday Brahem @odaybrahem]

Okay I saw someone posting this showing that kimi is claiming its claude so I tried it myself, wild!

[Oday Brahem @odaybrahem]

@Jaytel Love this!