quanto
Is World Labs' Atlas genuinely novel? Their demos do not seem to be better than the existing state of the art.
Fei Fei Li has been criticized even in the ImageNet days as a shower than a doer. Her roadshow past two years extolling "world model" in vague terms didn't show much insight, and her company's demo seems to merely rehash what has been available in the field for years. If you haven't seen the current SoTA in Gaussian splats, then World Labs' demos may seem cool, but they are mostly standard in the field now. If someone is more familiar with the inner workings of World Labs, I am happy to be corrected (but inner details wouldn't change the fact that demos are no better than SoTA).
I really had high hopes for World Labs, but I am quite saddened to realize that perhaps all this was just a financial maneuver and the detractors were right from the beginning.
source: I used to be in the downstream field: Gaussian spats for robotics. The exact field that World Labs' is supposed to help.
A minor correction: both the original Cheetah and Mini Cheetah motor designs were QDDs and were using BLDCs. The innovation was from using a off-the-shelf BLDC model as much as possible instead of a custom-design/modify one. Moreover the placement of electronics and gearbox was much more compact, leading to the modern day design. If any thing, the original Cheetah motors were actually simpler than those in Mini Cheetah.
Wouldn't it be great if Iran becomes our friend and trade partner in two decades?
Then, the question becomes why is the B-2 bomber (and latest stealth fighter jets) not angular like F 117.
Scientific computing has come a long way since the days of F-117. We do not need flat surfaces to scatter the beams away. Finding optimal curved surfaces that scatter the beams away is just a harder math problem.
If I remember from my numerical analysis days, it was the ease of numerical computing that necessitated the facets. Think 3D graphics from the 90s. All polygons.
Is there a typo somewhere?
David Ha, CEO and co-founder, was one of the youngest managing director at Goldman Sachs before doing ML at Google. His ML publications were considered top-notch almost a decade ago. I had high hopes for him when he raised money and founded Sakana.
I do agree with some comments here that perhaps this particular product is not well thought out. I also agree with the criticism that David calls Sakana a frontier AI lab while making money just selling AI B2B applications to Japanese businesses. I also agree with the assessment that Sakana has abrasive and antagonistic, sometimes openly hostile, recruiting tactics. I also agree that his then-impressive publications may have lost their luster in the age of LLMs.
However, the man is clearly driven; and he and his team may have more to offer in future. I admire the man for not taking the conventional AI-research career path.
The colleague claimed that there is no special magic. It's not that ASML is using some otherwise unknown laws of physics nor is any single step or component particularly special or novel. It's just that they meticulously optimized each step, and the sum of such steps is the winning solution.
In fact, this is probably why it's so hard to copy ASML. If there was a single magic component, a single or few engineers could be poached away to a competitor to copy it. However, copying a well-optimized company with many simultaneous optima is a much harder task.
Our discussion was in the context of why our quant hedgefund competitor was performing so well, far above the market norm. By nature and design, quant finance is an incredibly efficient field (and most techniques are more or less known by veterans), and we had thought unlikely that one fund could do so much better. Our conclusion was that this fund must be the well-optimized ASML of our field. My colleague happened to know the founder and indeed that was his personal impression as well.
Does the phrasing imply that Google sees itself as #1 and Nvidia as a close #2 that may rival its greatness? not to mention other big contenders.
input: I just ate lunch.
output: I’m thrilled to share that I’ve just successfully completed my midday refueling session!
Nutrition is the ultimate foundation for high-performance leadership and sustained productivity. Taking this intentional break has allowed me to sharpen my focus and realign my goals for the afternoon sprint.
How are you optimizing your energy levels today to drive impact? Let’s connect and discuss the power of strategic self-care in the workplace.
#Leadership #Productivity #HighPerformance #WellnessAtWork #GrowthMindset
I do some tangent work from this field for applications in robotics, and I would consider (metric) depth estimation (and 3D reconstruction) starting to be solved only by 2025 thanks to a few select labs.
Car vision has some domain specificity (high similarity images from adjacent timestamps, relatively simpler priors, etc) that helps, indeed.
True for some approaches, but numerical analysis does account for machine epsilon and truncation errors.
I am aware that Inria works with Coq as your link shows. However, the link itself does not answer my question. As a concrete example, how would you prove an implementation of a Kalman filter is correct?
My question is the same, albeit more technically refined. How do you prove the correctness of a numerical algorithm (operating on a quantized continuum) using type-theoretic/category-theoretic tools like theorem provers like Coq? There are documented tragedies where numerical rounding error of the control logic of a missile costed lives. I have proved mathematical theorems before (Curry-Howard!) but they were mathematical object driven (e.g. sets, groups) not continuous numbers.
I wonder if AI's compute graph would benefit from a language-level rigor as of Coq.
This is incredible. It's far less than I would imagine. It represents how well optimized the warehouses are. If we roughly estimate a median product price to be $20, then the automation represents less than 2% cost saving. Of course, Amazon is at a scale that this is still net positive despite all the R&D cost. But if automation was to reduce the cost of living, there are probably better areas to focus on.
What's interesting is that action tokens are learned from video. In other words, the training dataset does not include actions like "go left" and "go right"; and these actions are learned from the pixels that moved. This means that learned actions may not map exactly to the game actions available to the user. That means we (humans) cannot necessarily use this world model to play the game.
I suspect the inferred actions probably directly correspond to human-understandable actions; and after playing with the action tokens, a reasonable human can probably guess what, say, the third action token in the dictionary corresponds to ("jump"). This is likely as game actions are sparse (in both time and action spaces) and often independent/orthogonal (in action space).