
🎙️ AI reactions, explanations, and all-around good times with @braunch @eighttwo_three and @dylan_halliu
Latent Space
Joined July 2024
- Tweets222
- Following4
- Followers410
- Likes39
The Feature Crew retweeted
I wonder if a post AGI world has a fast economy and a slow economy. In the vain of the famous book Thinking Fast and Slow.
The few corporations that survive, assuming some form of UBI or dividend, will compete in the fast economy and most of us will watch it like a sports team. Cheering on victories and lamenting defeats (setbacks on some big project that impact our dividend). We may feel a collective but distant relationship to that economy.
Those dividends are then circulated in the slow economy. Local, human goods and experiences that are decided offline. I expect local arts, 3rd places and farmers markets to thrive in this economy.
Agents becoming the majority of life online will make much of it alien. That’ll be fun to visit and I’m sure looking forward to a holodeck but it also makes it untrustworthy. You’ll never know if something online is human. That friction will cause a shift to the local.
If you assume a location independent dividend that’ll mean people moving and reorienting their lives around choosing communities that are local, in person, decidedly messy and human.
Today for many it seems a local experience isn’t real unless validated by an online audience. An AI first internet reverse that for many (perhaps most?).
The Feature Crew retweeted
Obsessed with this goolge-earth-like diegetic map concept
still needs some tweaking
The Feature Crew retweeted
I burned all the codex resets on a low poly sandbox game
then I asked it to record a montage in-engine:
The Feature Crew retweeted
Spent a few hours with 5.3 codex last night building this space game. TBH kinda floored:
- multi-planet solar system
- seamless planet-planet transit
- board-able space stations
- easily runs 4k 120fps on my MBP
All this in <10 prompts within a single 5hr codex quota
The Feature Crew retweeted
This was claude 4.5 opus ONE SHOT in vibeworlds
I need to figure out how to get this shipped ASAP without going broke
HMU if you know generous people at @AnthropicAI
The Feature Crew retweeted
Vibeworlds turns one week old today!
To celebrate, we are rolling out features that let you better steer world generation:
🖼️ Image to World
🔁 Remix worlds
Available now for all research preview users! 🔗⬇️
The Feature Crew retweeted
Been quietly working on this for a while.
Vibeworlds is an early step toward more personal, more immediate worlds - places you can actually step into moments after imagining them.
nitter.cf/gen_holography/status/…
The Feature Crew retweeted
Gemini 3 Pro is up on vibe bench. Not the jump I was hoping for - hopefully it is better in other domains.
The Feature Crew retweeted
A quick test of the @OpenAI new open source models: 120b and 20b MoE models. With the @FeatureCrewPod planet generation prompt
The comparison is to the best one I've seen - a yet unreleased OpenAI model, but then Sonnet 4, Kimi K2, Qwen 3 Coder and GLM-4.5 for comparison. The comparison is not really fair as 120b and 20b are far smaller than the others, but hopefully this will give you a a sense of where to place them. Other benchmarks are looking better than this one, so this shouldn't be taken as the last word on these models.
The Feature Crew retweeted
There is a CRITICAL bug on @ChatGPTapp (moblie) that I have now seen repro multiple times.
Issue: A/B feedback is inverted. The "preferred response" you select blown away and the other is saved.
If oai is PFT on the wrong responses, they are training the model to be worse
The Feature Crew retweeted
Man still thinking about @sacca’s point on Tim Ferris’ podcast that theatre might outlast in-theatre movie watching as AI creation ramps.
The Feature Crew retweeted
New paper alert - MemLong looks to be an advancement in long-context language modeling, extending context windows to 80k tokens. It surpasses state-of-the-art models in efficiency and performance, offering substantial improvements for AI tasks involving long texts. 🚀 Check out the paper here: arxiv.org/pdf/2408.16967
#AI #NLP #MachineLearning
The Feature Crew retweeted
We’ve been working on Copilot in the new Planner - check it out here
Try Copilot in Planner (preview) in the new Planner in Microsoft Teams today! With the power of generative AI, Copilot streamlines the planning, management, and execution of your work. Learn more: msft.it/6019lizOt











