nxtfari

Karpathy also. CS231n first came out in 2015.

Given that Google is the named partner, it’s almost certainly Gemini. Also would make sense, Gemini’s (only, lol) strength is search grounded responses.

I was so excited to read this, get an inside view of how this problem must actually be much more nuanced than it seems for some reason not observable from the outside, just to read 10k words of whining and pleas for sympathy. When he finally makes it to the technical section it’s just “it’s a lab culture, there’s a lot going on.” There were some extremely inexcusable mistakes made apparent to anyone who knows anything about security at all. If the lab is spinning off experiments faster than you can improve your security posture just say that and own up to it (it’s what he appears to be trying to say in the worlds most mealy-mouthed way).

What you say is true, and Moonshot definitely has an agreement with Fireworks, but it still just feels wrong. It’s not the America I grew up in, where if you took, you gave back. I know that’s not a very coherent and practical position, but it feels true to me.

The more I learn about Fireworks the more unsavory they seem as a company. I don’t care what the license says, Moonshot has been openly improving, sharing research, and providing weights for the models that make up your entire bottom line, and the moment you can improve them in reciprocal it’s closed weights, “this is our own proprietary” nonsense? Where are we that China has better open source ethos than America?

Haha this did help me generate empathy for the assistant. Neat idea.

They do have robotics, it’s just not as far along yet as you might expect looking at their progress in other domains. Look at Figure and Gemini Robotics for good examples of SOTA. They’re impressive, but not ready to be in the streets quite yet, I would guess one more year.

> Privacy preservation: an outside observer cannot determine whether any pair of reference images were taken by the same device. Image contents are not exposed to Apple or anyone else.

I am curious if the PCC processing makes it resilient to the method from a couple years ago of fingerprinting the microdistortions in an individual phone lens to tell when two photos were taken by the same phone.

We’re reaching quadratic slop. Slop projects that don’t understand what they’re shipping built on top of slop projects that also don’t understand what they’re shipping. Magnificent.

Good feature but god I hate there’s a whole new keyword for a minor modification on an existing concept. Very unnecessary.

As the saying goes every 5 minutes in Silicon Valley another company reinvents ROS

Agree, I remember when even half precision made its way into C# sometime around 2020 (I didn’t know much about ML then) and I thought, well I guess that’s a worthwhile tradeoff but I can’t imagine going lower. Lo and behold (1-bit Bonsai) how much lower you could go.

Sure, then you should allow that they are also tools for transforming sequences of questions into answers. Language models are based on compression of information, using them as a knowledge base is entirely within capability.

Surprisingly incoherent for Anthropic and Dario (cue peanut gallery — “always has been!” No, I don’t think so. I think this is new).

It seems to me like there is just no good answer to how one could possibly stop open weight models from being used for nefarious purposes. How are you going to enforce guardrails on open source? The only way is to turn the USA into a 1984-type totalitarian surveillance state (even more so than it is). Unable to say that, we just get this floundering instead. How long is not giving them chips going to slow them down? Until we RSI? Then what? Just because RSI runs off the exponential doesn’t mean that the eventual open-weight Moonshot Mythos won’t be able to make bioweapons. Genuinely what is the endgame.

They must have done the math to show even if you peg Kimi K3 generating 24 hours a day for an entire month it doesn’t exceed 10k in opex? Very interesting.

Jagged frontier is not the same as being benchmaxxed. Benchmaxxed is à la Goodhart's Law "when a measure becomes a target, it ceases to be a good measure." Jagged frontier is about how models that seem superhumanly intelligent at one category of tasks (e.g. coding web applications" can seem toddler level or worse at another category (spatial reasoning) because the training corpus doesn't generalize to there.

Haha, this is awesome. Hadn't heard of it, thank you for the link.

If you haven’t been really running and testing these models yourself, they are all benchmaxxed. No matter how close they score to frontier on whatever metric, they always fall apart in real world tasks and their token efficiency is ridiculously bad. Fireworks has incredible incentive to make this claim in a headline, because Fireworks hosting K3 for you is pure profit for them, unlike when they host closed source models.

Very high quality and info-dense article. Cheers.

This is really smart, like the author said, old idea but cleverly applied.

In case anyone wants a summary: don’t one shot, don’t use plan mode and hand off the plan to cheap executors, ask the frontier model to explore, create a todo list, and then start when it feels confident; stop it after first code edit, then prefill the context to cheap executor to continue.

Everyone using Claude Fable to verify this proof is so funny. If you read the definition of the Jacobian Conjecture and (I am not exaggerating this) have passed a college Calc 3 class, you can just verify the proof yourself in 30 seconds. The problem was very hard to solve but the counterexample is very easy to verify!

--- edit, adding an explanation:

To summarize it, the conjecture says if you have any multi-variable polynomial function that maps an input to an output in the same dimensional space (take for example: F = (x+2, y+2), which maps 2D space into another 2D space), AND that function has a constant-valued non-zero Jacobian determinant, THEN the conjecture is that the polynomial has an inverse, meaning basically you can find a polynomial that turns the output space back into the input space.

Fable provided the example polynomial (which was very hard to do) and the coordinates which if you plug into it, results in two points being mapped to the same output point. This means that the polynomial can't be inverted, because if you have that output point, how do you know which input point it came from?

You can just plug in the two coordinates it gave into the equation and verify that you get the same output point from both. That's the contradiction of the conjecture and it takes 30 seconds.

---

Something something outsourcing of thinking something.

Eh, Minimax M2.7 also took a similar amount of time (actually longer) between availability and weights release.

you could live (100k a year) with the interest of 0.004% of that in the bank.

Apple (and the ARM ecosystem as a whole) has never really needed massive GPU compute before, it’s always been about power efficiency and just enough GPU oomph to make UI fluid. Even historic Mac Pro workloads never really needed tons of GPU prowess, the heaviest power users were primarily taxing video encode/decode and 2D raster effects (Adobe Suite etc) so that’s what they focused on. By contrast NVIDIA’s entire game has been raw GPU power for decades. M1 Max was really prescient in hindsight and has set the stage for, all things considered, Apple to be not as far behind as it could have been today.

I think we should make it illegal to not specify the quantization in the headline for these types of posts.

this is really cool but it seems very unlikely that someone targeting an exotic system not supported by rust (mostly embedded and ancient mainframe targets) would be willing to trust a beta transpiler to not inject any bugs or leaks in the process of turning rust to c. nevertheless, very cool.

An important idea I’ve observed true across industries is “prices rise like a rocket and fall like a feather,” meaning that even though price rises are usually genuinely driven, you can bet that once they’re up the involved parties are doing whatever they can to keep them up. Humans are greedy.

More formal reading: https://en.wikipedia.org/wiki/Asymmetric_price_transmission

this is really clever, props

Location: NYC / LA

Remote: yes

Willing to relocate: NYC or LA

Technologies: Motion planning, autonomy, embedded microcontrollers, embedded Linux, perception systems, controls. C, C++, Python, Rust

Résumé/CV: https://github.com/nabeelsherazi/nabeelsherazi.github.io/blo...

Email: see resume

7 YOE in shipping robots to customers, lots of safety critical stuff, looking to move into urban autonomy (self driving cars, delivery bots, construction bots, humanoids, etc). Looking for a startup with a mission I can get behind.

they’re doing m7 on the intel 18a fab, which is exactly that