> remember how they mentioned they built multiple Codex prototypes, it must've sucked to see some other people's version chosen instead of your own
Well it depends on people’s mindset. It’s like doing a hackathon and not winning.
Most people still leave inspired by what they have seen other people building, and can’t wait to do it again.
…but of course not everybody likes to go to hackathons
> OpenAI is perhaps the most frighteningly ambitious org I've ever seen.
That kind of ambition feels like the result of Bill Gates pushing Altman to the limit and Altman rising to the challenge. The famous "Gates demo" during the GPT‑2 days comes to mind.
Having said that, the entire article reads more like a puff piece than an honest reflection.
Uhm. This seems more of a case of slow and expensive, can we at least hope it’s good?
The plane was initially commissioned in 2018:
- originally planned for delivery in 2024, the first aircraft’s timeline has now slipped to at least 2029, with further delays possible;
- The fixed-price contract negotiated under the (first!) Trump administration capped costs at $3.9 billion, but Boeing is already $2.5 billion over budget
Can you guys explain what this would be bad for the OpenAI and Anthropic of the world?
Wasn't the story always outlined to be we build better and better models, then we eventually get to AGI, AGI works on building better and better models even faster, and we eventually get to super AGI, which can work on building better and better models even faster...
Isn't "super-optimization"(in the widest sense) what we expect to happen in the long run?
First of all, we need to just stop talking about AGI and Superintelligence. It's a total distraction from the actual value that has already been created by AI/ML over the years and will continue to be created.
That said, you have to distinguish between "good for the field of AI, the AI industry overall, and users of AI" from "good for a couple of companies that want to be the sole provider of SOTA models and extract maximum value from everyone else to drive their own equity valuations to the moon". Deepseek is positive for the former and negative for the latter.
I believe in general the business model of building frontier models has not been fully baked out yet. Lets ignore the thought of AGI and just say models do continue to improve. In OpenAIs case they have raised lots of capital in the hopes of dominating the market. That capital pegged them at a valuation. Now you have a company with ~100 employees and supposedly a lot less capital come in a get close to OpenAIs current leading model. It has the potential to pop their balloon massively.
By releasing a lot of it opensource everyone has their hands on it. Opens the door to new companies.
Or a simple mental model, there has been this ability for third parties to get quite close to leading frontier models. The leading frontier models takes hundreds of millions of dollars and if someone is able to copy it within a years time for significantly less capital, its going to be hard game of cat and mouse.
For years, I was an Apple Watch user: I assumed that all medium/top end trackers were the same, and that Apple Watch was pretty much the benchmark.
…but now that is have had the opportunity to use extensively Garmin watches, my experience is that they offer far superior accuracy, precision and technical details for activity tracking and sleeping than Apple Watch.
My picks would be in the following order:
1) Garmin high-end watches, they are truly a work of love
2) Aura ring, because of great convenience and reliability
3) Apple Watch, because they are great all-rounders
4) Coros, Suunto, Whoops, because they are highly reliable, but lack some of the smart functions
5) Withings, Fitbit, etc…, they are a solid option, but they generally lack distinctive features/capabilities
I would stay away from any brands offering super cheap products, due to privacy concerns and lower reliability and lack of advanced features.
I would love to hear more about your opinions on this as someone who has been experimenting with as a semi-serious runner after years of being Apple Watch exclusive.
What I see is benefits around battery life, form factor (buttons are awesome), and good native support for "compound metrics" like Endurance Score, Hill Score, Training Status, etc.
But when it comes to actual stats and metrics, Apple Watch feels superior in most ways. Garmin sleep tracking anecdotally feels much less accurate. It baffles me that it only shows pace to the nearest 5 seconds during a workout. It confuses me that it only shows a Vo2max estimate to zero decimal places.
Then, Apple Watch is at least 10x more customizable via third party apps. Want a Whoop-like experience with strain score, recovery score, etc.? Bevel and Athlytic are there. Want a much more in-depth and customizable workout experience? WorkOutdoors puts Garmin to shame here.
What am I missing that makes Garmin so pervasive, while Apple Watch is derided as "not a serious sports watch"?
I'm hardly a serious runner, but I'd say the pros you laid out for Garmin are quite nice, and the cons are inconsequential to your average fitness tracker user. I'd probably argue they're inconsequential to everyone but the absolute elite and, for them, are pointless.
Sleep tracking is hard to action on for the average user outside "you slept this long" and none of the writst-based devices are that good anyway.
Pace to sub 5 is a little more annoying, but probably not useful for the majority considering most people are just running, not craning over their watch the whole time.
VO2 max is also a wild estimate, and I'd hazard it's not particularly accurate for the average person. It's off by close to 20% for me, and I should be a pretty good candidate.
On the flipside, you can get tons of data out of a Garmin that costs significantly less than an Apple watch. Plus, the majority of Garmins sold are fitness devices with some smart features, with Apple watches being primarily a smart watch. While maybe not justified (I think the Apple watch features are quite nice) I'd expect that's a major part of the reason Garmin has the rep it does.
If someone is buying a device to run, most would recommend the cheaper, light, simple, specialized, long battery life watch over the opposite. If you already have an Apple watch, it's probably a no brainer.
For the high-end Garmin devices, it's a little more complex, but not many people are considering a US$800+ device without knowing the nuances of the discussion, or having enough money to not care.
I do think the pace having more granularity than five seconds is important for anyone who's doing any kind of speed work, where a pace off by 5 seconds can result in a fairly significant variance. Admittedly I am not a total novice, but my 5k and 10k pace times are about 10 seconds apart, and I do some interval workouts at 5k pace and some at 10k pace. 5 second granularity doesn't give much wiggle room there! Although of course, GPS and cadence-based paces are also estimates, so maybe the 5 second accuracy is better than 1 second which could inpsire a false sense of confidence in the estimate.
As far as Vo2Max goes, totally agree – my lab test results vary widely from both watches. However, I think that actually makes Apple's 1 decimal place more significant – it has a lot of value in offering a fitness trend, even if it's inaccurate. I might train hard for 3 weeks and see 0 movement in my Garmin Vo2Max, whereas I might see a 0.3 increase in the Apple Watch. This is valuable for even the novice runner.
I feel the important piece to remember with VO2Max estimation is: Its an estimation. Its significant figures [1]; reporting the value to one or more decimal places communicates a level of confidence inappropriate for how inaccurate these estimations generally are. Especially the Apple Watch's; Garmin's is known for being pretty decent, usually +/- 2, but Apple Watch's is all over the place and is infamous for being really inaccurate.
Clamping pace to 5 seconds is a similar idea. GPS isn't super accurate: within 16 feet some sources say [2], though it gets better if you've got dual band, if you're moving; but it gets worse when you don't have an open sky. Just ten feet of GPS inaccuracy over a ten minute mile means your recorded pace is somewhere between 9:58/mile to 10:02/mile. And, experimentally, these systems are way, way more inaccurate than that: on a recent bike ride, with no major sky obstructions, I wore both an Apple Watch Ultra 2 and Garmin Enduro 3; the AWU2 recorded 25.05 miles, Enduro 3 recorded 25.18 miles. That's a difference of ~686 feet; ~27 feet/mile.
That's very true, and I'd love to see some actual documentation on how they get to pace numbers.
I'm in the same boat with regard to 5K/10K pace, but I reckon it's probably not a huge issue in the long run. While plans specify those times, I think it's more about shorthand for effort zone where 5K is "this hard" and 10K is "a little bit less hard".
VO2 max improvement is a good point, though, and I'd probably agree. If I had a hazard a guess, Garmin would say that their training productivity tracker/race estimated are the preferred way of presenting that data. as an aside, I think VO2max has sorta been coopted as a "fitness number" when it actually represents a very specific thing that may or may not be emblematic of actual performance in the majority of cases. It is nice to have a a single value to look at that can sum up whether what you've been doing lately is productive, though.
That could just be me coming from the world of cycling where watts are king and there's far less variability. In my mind, all these running stats are mushy, but that might not actually be the case.
If vo2max is displayed without decimals it would take months to see progress for most people starting running. It’s baffling that they would make such a mistake.
I was considering a Garmin watch, but if they make such a stupid decision regarding vo2max then what other mistakes are lurking in their apps?
IMO the biggest reason why the Apple Watch is oftentimes interpreted as an "unserious" exercise smartwatch is actually quite simple: The display & lack of physical buttons makes it difficult to interface with in the variety of conditions that outdoor activity enthusiasts often find themselves in. If its bright out, the mps displays on many Garmins will outperform OLED. If its raining; good luck using a touchscreen. If you're wearing gloves; ditto. If you've just ran a marathon, you're dying, your vision is blurry, you're sweaty and collapsing, that "swipe over a screen then click the end workout button" workflow is the end of the world; it wasn't designed by someone who has ever been in that situation, its designed for and by people who take their nice little walks to the cute little grocery story.
Battery is another less major factor: Even the AW Ultra 2 struggles to make it through a full marathon run (~70-90% battery usage IME) and that's not an uncommon-enough situation for users of the quote"ULTRA"endquote to be an invalid criticism.
Nothing else matters. Your comment continues into talking about sleep tracking and recovery scores and strain scores and third party apps and literally none of that matters. That's silicon valley brain stuff that many customers don't care about. The Apple Watch is, to some people, a Bugatti without a steering wheel; it gets a lot of the basics wrong.
One note though: Many Garmin users would also say that Garmin is, sadly, also losing track of what their core userbase wants, as the experience has become more buggy and less focused over the years. I'm not asserting that Garmin is king and Apple are idiots; Garmin just has momentum and is generally great at the things its users care about.
I actually can't imagine what it will do for video gaming. Maybe enhancing cut-scenes, but then why can't they just do the performance and rendering using the gaming engine in realtime?
Perhaps AI will help with procedural generation of environmental details within a pre-built game world. This way the AI isn't burdened with generating the whole scene, but only the clutter of objects and textures - things that usually take a long time to build by hand.
For example in Train or Truck Simulators, I see examples where someone has put effort into making that farmhouse in the distance nicely detailed, but other times it's just a simple structure. If AI were tasked with "distant details", the whole game could look more polished.
The photo app redesign looks great on paper, but is terrible in practice.
Watching videos in full screen, with audio on, and without controls overlay (I hope I am not asking too much!) requires multiple taps and several frustrating attempts every-single-time.
Activity on your shared albums has been relegated to a small text-only link, and you only get to see small photos, and it doesn’t smartly tiles photos together anymore (photos are now randomly chopped to fit in the frame).
The photo app gets badges, but then you don’t know where to find the new activity inside the photo app.
There is no way changes like these were created, tested, and validated by a well-functioning team
Also, watching videos full screen and scrubbing around in the video requires backing out to the smaller view. Then you get to fast forward/rewind. Then you have to tap to watch full screen again. Everything about this experience has gotten worse. I can’t even come up with a use case in which this is an improvement.
While you're technically correct, the parent is more correct. Competitors have to pay normal launch rates. The competitive service needs to include those costs to end users.
Starlink "pays" for launches at cost. While we don't know what SpaceX's cost margins are, they are not trivial. To setup a low orbit constellation is extremely expensive and competitors lose millions per launch that Starlink gets to reinvest.
There's been 136 launches of Falcon 9 for Starlink. ~US$62m per launch? If their margins are 20% that's that's $1.6b in savings. And I bet F9's margins are closer to 50% - supporting Starship and more.
I haven’t decided yet where I stand on this, however…
…As I understand it, the intent of the law is to avoid that large companies book billions of revenues in the country, but then use price-transfer schemes to book the profits somewhere else, and not pay any taxes locally.
The net effect of the law on larger internet companies is probably marginal, but I am not sure the law was designed well enough, so as to avoid collateral damage on smaller internet companies
The way Italy is implementing their own law looks kinda like ultimatum for others to implement minimum corporate tax law. As was agreed, but then never done because China and US are arguing details and everyone knows last one to implement theirs is likely a winner economically.
You can sit around and wait for decades until this might get meaningfully implemented, loopholes get closed and enforcement is strict. Great step by Italy to refuse that and just implement a measure in the meantime. Can always be repealed if indeed things get fixed on a global level.
I agree. Also, especially in the extremely crowded and noisy context - what would have been the chances to have the demo working so well?
In fact, even if the robots worked very well autonomously, you would still have wanted a way to ensure that the demo is successful - the same way Steve Jobs did with the iPhone demo, Larry Ellison did with the Oracle servers demo, etc.
So many stories like that in the history of famous product launches.
The one thing that bothers me a little is that if you look at the robots dancing, they are only moving the upper body; their feet are always on the ground. I would have liked to see them having enough ability to dance and move the legs too… then, again, maybe the gazebo they were in was just too space-constrained, or it was just too risky to do that in the demo - given the crowd, and all the chaotic party context. When you set up a demo, you have to account for the edge cases where your product glitches, not just for what it mostly does very well.
Anyhow, these are all AI issues (as opposed to mechanical ones), and, at the pace AI is evolving, it is not hard to see how these types of issues get ironed out over the time horizon leading to the launch.
The Optimus demo did do a great job at actually making people see a world in which robots just roam around and interact with humans everywhere. .
> these are all AI issues (as opposed to mechanical ones)
Actuation is still a massive problem in humanoid robotics. We have over 650 muscles. A humanoid today can't even approximate that. Sure, a robot might not need that many actuators to be extremely useful. However, to be general enough to be able to interact with any human environment, the number of required actuators will not be trivial.
Add to that gearing, couplings, driver electronics, encoders, thermal management, calibration, noise, maintenance and other per-actuator requirements and the picture quickly becomes overwhelming.
This is an area that is still looking for a significant breakthrough.
- Small, powerful 3-phase servomotors are cheap and easy to obtain. Mass production of drone motors has advanced small motor technology considerably. Tiny motors use to be either toy-grade junk or expensive Swiss precision. That's improved.
- Motors with built-in encoders are, at last, available. Encoders used to be fragile plastic boxes stuck on the end of the motor. Also, thermal sensing inside the motor is common, so you can tell if you're overheating it.
- Permanent magnets are small and powerful, and have such high coercitivity that you don't have to worry about demagnetizing them if you over-drive the motor. The main limit on motor power is cooling. You can way overdrive a motor momentarily, like muscles.
- Motor controllers are now small and cheap, They cost about $1000 per motor two decades ago. The power semiconductors are small. Controllers can be programmed to use very high power levels briefly, monitoring thermal sensors.
It would be nice to have good linear actuators. Linear motors do exist, but never really became a big thing.
I'm totally with you on the evolution of motor tech because of drone and also personal mobility (scooters and hoverboard motors are a steal for what they can handle).
While high torque motors got way cheaper, especially with MIT Cheetah "clones" getting easily available, they're still at least 200-500 a pop (depending on the torque needed for each articulation) from what I could find.
I might not know where to search for the real gems though. Where do you search for cheap powerful servomotors?
It's impressive how well the hardware seems to work now, though the software is still clunky. You can see how well the hardware works under human nervous control in the recent MIT bionic foot https://youtu.be/1tD7qd68i3o?t=36
Well it depends on people’s mindset. It’s like doing a hackathon and not winning. Most people still leave inspired by what they have seen other people building, and can’t wait to do it again.
…but of course not everybody likes to go to hackathons
reply