One benefit answer.ai has over some of those labs is that it's building software, so expenses are mostly just salaries and GPU time - no need for massive capex on fabrication equipment or anything like that. And since we're focused on efficient fine-tuning and inference rather than foundation model training, even the GPU spend is likely to be extremely modest compared to some other players :)
Nice work - I get ChatGPT to write ffmpeg commands quire frequently, this is a logical next step :) One thing: you may want to clarify that "Upload" also does the processing with whatever text is currently entered. I expected to upload, then edit the instruction, then run. Maybe renaming the button to "Upload and process"?
I tested this out on a topic I'd been discussing with some fellow researchers, and it pulled in the papers we'd chatted about plus a bunch of related ones that look very relevant and interesting. Congrats on a cool project!
FWIW I was able to get about 20% accuracy (perfect 4/4 groups) with ~50% of groups correct on average and most mistakes being groups with 3/4 right (so at least on the right track) with my first attempt at 0-shot prompting. The prompt goes something like this:
```
You are playing [game info] with [word list]
Follow these steps:
- consider possible groupings as initial brainstorming (>4 groups)
- propose a first hypothesis based on the likely-looking groups
- reflect on whether that grouping works
- revise if needed then submit the final predictions
```
Having it start with word one of group one as the first output token seems unlikely to work from my intuition about what these models can do. Heck, I can't solve it that way! Burning some tokens on exploration and hypothesis building leaves it with the easier task of choosing plausible groups from the proposed options. System 2 thinking vs system 1 perhaps.
I tried with clustering similar embeddings but it did extremely poorly (~0%) since the groupings are often deceiving with words in a group only having one small way in which they're connected and lots of spurious fake groups to throw you off. Maybe looking for groups with high similarity on only a sibset of embedding dimensions might help, but I didn't have much time to play either :) A notebook to get you going if you do want to play: https://colab.research.google.com/drive/1KJeSB9Q5XzSeT9ONUJ_...
3D printers are great general motion platforms! I'm a big fan of things like this, which re-purpose cheap printers in new ways. I've had fun adding a pen to mine to turn it into a plotter, or strapping on a camera with a high-magnification lens for stacking images of small objects.
One reason this implementation is exciting is that the code that accompanies the Gaussian Splatting paper requires companies wanting to use the tech commercially to pay for a commercial license. There are existing viewer implementations, but this seems like the first alternative that would let me do everything without depending on the Inria code. The code being nice and clean is a bonus too :)
Garbage in garbage out still applies (I tested with some shaky phone video and the results were a mess) but when done right this looks incredibly promising for future VR, film and gaming applications.
Gaussian Splatting is a relatively new technique, somewhere between a NeRF and traditional photogrammetry, which allows for fast rendering of 3D scenes captured from 2D input images. Luma labs are rolling out new features based on this technology including fast interactive viewers for WebGL and metal. The scans look incredibly high-quality.
For anyone curious to learn more about the incredible lives of seabirds I can highly recommend the book "The Seabird's Cry" by Adam Nicholson. Poignant and poetic, with each chapter covering a different species, their amazing adaptations, and more often than not the new challenges facing their survival thanks to human activities.
Definitely. One of the time-consuming parts of rendering is sorting the gaussians by distance to the camera, which for two nearby cameras could be optimized. This also goes for adjacent frames - assuming smooth motion im pretty sure there is some speedup to be had by assuming the previous sort will be close to correct rather than starting from scratch each frame.
This April Fools joke got me back in 2013, and my family have never let me forget it! I should have checked the article date...
Thankfully our microwave started working again once it cooled off and all that was hurt was my pride.
Cloud GPU providers are running low on capacity at the moment as people frantically suck up capacity to hop on the AI bandwagon, raising worries about availability. So having guaranteed access is maybe one motivation for local GPUs. But for me the main reason to go local is more psychological. I've mostly used cloud compute up until now but whenever I'm paying an hourly cost (even a small one) there is a pressure to 'make it worthwhile' and I feel guilty when the GPU is sitting idle. This disincentivizes playing and experimentation, whereas when you can run things locally there is almost no friction for quickly trying something out.
I tried to test this with nanoGPT in an afternoon, since the code change is pretty minimal. It's hard to get conclusive results at that scale though - to be able to say anything with confidence you'd need to run multiple tests, figure out if the 'outliers' mentioned only appear above a certain scale, find good tests for quantization performance that work on small enough models that you can iterate quickly ... It's doable but still lots of work, enough that putting out the idea and hoping others with more time+compute will try it out seems a valid strategy to me :)
More generally though I definitely agree that the trend among 'improvements' to transformers has been things that don't turn out to work in practice.
Perfect illustration of how tricky it can be to draw a firm line.
Another survey in this same genre is "the rape spectrum" with >5k respondents ranking scenarios: https://aella.substack.com/p/the-rape-spectrum-survey-result...
Tough topics to discuss (I see why the OP went with vehicles in the park). I'm glad I'm not in the content moderation business!
I noticed an apparent shift recently (for the worse) using Bing in creative mode, which is also supposed to be GPT4. Shorter answers, much more work to get it to output code, and maybe more bugs in the code it does produce...
It's funny, I really did feel like I'd lost something when I noticed it!
Some things to note:
- This was finished last year, impressive now but even more so back then
- No code or models released BUT several authors have moved to StabilityAI and are working on their own improved open video models, which is hopeful as the field continues to move forward
- The paper uses existing image models as a base, and so a better base model (the new XL stable diffusion variant, or Midjourney's underlying model) will give even better results.
Really impressive to see such a large effort come together, with volunteers contributing code and data to make it work. Well done to all involved!
In their evaluation people preferred their model over GPT3.5 turbo about half the time. But since the data is open, as better base models become available others can train them with this data to improve even further.
Patiently waiting for someone to say "We trained LLaMA with extremely high weight decay, and find that performance is much worse. Due to licensing constraints, we are providing a version of delta weights that build on the original LLaMA weights..."
For normal people (non-programmers) computers are pretty terrible. I'm kind of curious what happens when everyone gets a Jarvis-like assistant that can do the scary stuff (open a terminal, type some commands, or possibly code up a nice UI in seconds) based on natural language requests.