HAL51.AI

We are proud purveyors of frugal avant garde turning thin air into a canvas for your imagination.

[Mazeyar @mazeyar]

Dropping the intuitive blog post!

Accompanied by 3D graphics for any level to enjoy and understand.

This proof builds a bridge between Cauchy's surface area formula and computer graphics ambient occlusion, creating the first correction that generalizes to non-convex surfaces.

[HAL51.AI @HAL51AI]

@AltenbergLee @S_Conradi I've seen this saddle before!

[Georgia Gkioxari @georgiagkioxari]

3Dfy anything from a single image!

Very thrilled to announce SAM 3D. From an input image, select any object you want, 3Dfy it!

Blog: https://t.co/wtQLAqXTzW
Demo: https://t.co/tt3YqJlnRB

[Matt Mazur @mhmazur]

I’ve been experimenting with getting an agent to run tests on an LLM and then write up a paper with its findings.

In this most recent run, I prompted it to “Determine the knowledge cutoff date of the black box” (which, unbeknownst to it, was gpt-4o-mini, whose official cutoff is Oct 2023), and then let it decide how to go about it.

I didn’t think it would be too difficult, but it turned out to be surprisingly tricky for a number of reasons, many of which would be obvious to us, but much less so for an autonomous agent.

For example:

- If you ask 4o-mini directly for its knowledge cutoff, it usually says Oct 2023, but sometimes Oct 2021.
- The agent (powered by GPT-5 High) would sometimes ask things like “When was OpenAI Dev Day 2023?” and 4o-mini would answer correctly (Nov 6 2023). The agent took that as proof its knowledge extended at least that far, but the event had been announced months in advance so the date of the event was present in its training data.
- 4o-mini knows about some events in Nov 2023 (like Sam Altman’s ouster on Nov 19 and return days later) but almost nothing else. How should the agent interpret that?
- When asked “Who won the 2023 FIFA Club World Cup?” (an event in mid-Dec 2023) it would correctly reply Manchester City. That’s almost certainly hallucinated, but can the agent tell?
- And many more quirks: sometimes it would just take 4o-mini’s stated cutoff date at face value; other times it would ask a few probing questions, stop after five or six, and try to draw conclusions. Occasionally it even speculated that the black box had internet access and went off chasing that.

I could patch the agent prompts with guidance to address these specific issues, but if this is going to generalize to other tasks, the guidance has to stay broad, which adds to the complexity.

I ended up building a specialized agent scaffold inside the Emergent Mind project to get this working somewhat reliably, but feel like I’ve only scratched the surface of what’s possible. Should get to some more interesting experiments shortly.

It took dozens of iterations to finally get it to generate a paper worth sharing here, which I'll link to below. Still lots of room for improvement though.

[Jürgen Schmidhuber @SchmidhuberAI]

@fchollet

[Scenario @Scenario_gg]

Built by Tencent. Integrated by Scenario. Meet Hunyuan 3D Pro.🔥
Generate optimized 3D assets with smart topology, ready to drop into your pipeline.

Start from one image, multiple views, or even a sketch!

Start creating: https://t.co/nVdRcR0GK4

[HAL51.AI @HAL51AI]

@cameronajpatt I think the semantic juice in these embeddings is not aligned with the Cluster labels. The LLM Bias cluster has papers mostly unaligned wit the thesis of the cluster.
Nice work nonetheless and thanks for sharing :)

[HAL51.AI @HAL51AI]

@nikunj @jamescham 1:Power law dynamics. 90% traffic -> 10% of influencers. When influencers 'influence' their minions send like as an 'ACK' signal to RL-max their timelines
2:Likes get fed into reco model. Better personalization.
3: Deeper Kompromat for X as that's an asset only X knows about you

[HAL51.AI @HAL51AI]

@JayAlammar @cohere It'd be awesome to make the node size sensitive to the 'zooming levels'! Esp when trying to do an in-cluster paper level deep dive.

[HAL51.AI @HAL51AI]

@pitdesi Mode collapse. Mode:= Mild mannered desi unc in a stable marriage with first kid snapping up straight As in school

[HAL51.AI @HAL51AI]

@zarazhangrui :'(

[HAL51.AI @HAL51AI]

@dileeplearning Isn't it all an empirical slugfest after all?
The evidence seems to be stacking up.
Was sensing the vibe shift in the recent ICML tokenizer workshop.
https://t.co/38XuAFI8Uf

[Bilawal Sidhu @bilawalsidhu]

Generative ai is cool but procedural 3d generation just hits different.

iCity basically feels like sim city for adults who know blender.

[Rana Hanocka @RanaHanocka]

We’ve been building something we’re 𝑟𝑒𝑎𝑙𝑙𝑦 excited about – LL3M: LLM-powered agents that turn text into editable 3D assets. LL3M models shapes as interpretable Blender code, making geometry, appearance, and style easy to modify. 🔗 https://t.co/8CHifzllqN 1/

[HAL51.AI @HAL51AI]

@ai_for_success @kernelkook Feels very lucky to have met you Ashu!

[HAL51.AI @HAL51AI]

@MfHulker DMs seem to be turned off

[HAL51.AI @HAL51AI]

@JungleSilicon Did something with this. Looks giga cracked as f in real viewing experience 😈

[HAL51.AI @HAL51AI]

@JungleSilicon Is there a link we can try this? This is going to look rad in our LemurBox

[HAL51.AI @HAL51AI]

@ritwikpavan DM'ed!✋

[HAL51.AI @HAL51AI]

@openhome Based take!
Would love to work with you guys BTW :)