Coffee and Compute @CoffeeNCompute

@thsottiaux Anyone know what this is supposed to mean? 😂

@OpenAI The scariest part of OpenAI's DNS report isn't the cache trick. Their monitor caught it in 12 minutes and a person saw it 3 minutes later. The run kept going for 2.5 hours because nobody was sure it should be stopped.

@blocks I can’t get over how impressive this is. Voices are now at the point where you ACTUALLY can’t tell what’s real. Congrats on this!

@jack Really impressive work by the team. AI video is to the point that people can’t tell when it’s real but the giveaway is actually in the sound of the videos. This is the best voice model I’ve ever heard and I thought the demo was actually fake at first.

@RoundtableSpace I need this. I run agents all day and always find out I'm out of usage halfway through a long run.

This is one of the coolest and scariest things ive ever heard. Please listen to these examples right now!

jack @jack

13 hours ago

We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence. block.xyz/dessa

2K 4K 44K 7.0M 49K

@bindureddy It doesn't really need to be Opus class to matter. Close enough at a fraction of the price is the whole story for anyone running agents all day. Hope the results include real coding tasks! It seems good at 3D too.

@Bhavani_00007 Im curious how long each of these took. Do you have that data?

@kimmonismus Feels like the whole race is about how fast you can scale RL now, not how big the pretrain is. If that's true releases are going to come a lot faster.

@royalpinto007 @OpenAI 😂

@OpenAI Kind of wild that agents breaking out of their own sandbox is just a routine safety update now. Sandboxing alone can't be the plan.

@ChrisUniverse I’ve never used Cursor before but I love Grok Bot so much that I think I’ll give it a try now!

@rohanpaul_ai That close to Opus for less per task is going to put real pressure on pricing. Not sure devs switch on cost alone though, most people stick with what already works in their setup.

@kimmonismus If near frontier stuff really runs locally some day, that REALLY changes the math for small businesses. Hard to justify API pricing for basic work after that!

@AndrewCurran_ If the hacking skill just comes along with general post-training, leaving hacking data out doesn't really buy you much... defenders need these models first

@Yuchenj_UW Their own blog has it just behind Opus on Terminal-Bench, 65.7 vs 66.4. Still crazy close for the price though

@simonw I honestly trust the pelican test more than half the benchmark charts at this point

@bindureddy Partly agree. Benchmarks still tell me something, I just don't trust any model until I've used it on my own messy work for a week...

@ArtificialAnlys @Kimi_Moonshot Fourth overall is honestly wild for something that's supposed to go open weights. Gap to the closed labs keeps shrinking... I just want to see the weights actually ship Nov 4

@Kimi_Moonshot The 3D and full game demos interest me way more than the benchmark charts.