For ARC v1 it was found that it was much less resistant to brute force than intended/designed. This was improved in v2, which LLMs are currently doing less good at. Note also that ARC tasks are explicitly designed to be slightly-out-of-reach, things that are quite simple for humans, but current models are pretty bad at - designed to measure and enable progress.
But yeah there are many interesting approaches, and ARC is interesting to follow both because it attempts to measure ability to adapt to new takas ("fluid intelligence"), and because we have not saturated it yet.
Advertising is in many market like a tax or tariff - something all businesses needs to pay. Think of selling consumer goods online - you need ads on social media to bring in customers. Spending 10% on ads as COGS is a no brainer. 20% too. Maybe it could go as high as 50% - if the companies do not really have an alternative, and all the competitors ard doing it too? They are just going to pass the bill to the consumer anyway...
If you started with a deep neural network, one can't really use pruning to go all the way down to a parameter count that is directly intepretable (say under 100). One would at least have to try some techniques to get more disentangled representations. But local surrogate models are popular for explainability, see Shap and LIME.
For interpretable time series I would encourage to construct features and transformations the old fashioned way, and then learn it all end to end as a differentiable program. Then you can get the best of both worlds.
ESP32 or RP2 based boards with for example MicroPython/CircuitPython, or platform.io + VSCode. Though the good old Arduino IDE seems to be unaffected by this change though.
Adafruit also created and maintains CircuitPython, which is targeted very much at the same audience as Arduino. Very beginner friendly, great for quick prototyping/one-offs, but serious enough that one can ship small scale projects with it.
And Python is a much better language for that than C++ (which is what Arduino users do not really realize they are using).
For more on this perspective, see the paper On the measure of intelligence (F. Chollet, 2019). And more recently, the ARC challenge/benchmarks, which are early attempts at using this kind of definition in practice to improve current systems.
A RPi 4/5 has that? The OMAP 3/4 chips in Beaglebone is another alternative. Or Rockchip/Allwinner SoCs. Those are all Cortex A series though, but that can still be programmed bare metal if one wants. There are also Cortex M microcontrollers with support for SDRAM, like STM32H7. Or one can get an FPGA with DDR support like ECP5.
It i possible to have a digital twin which simulates the real physics of the robot and gives instant feedback based on that. It might help a bit to make the teleoperator work within the capabilities of the robot, be more in sync.
But is peobsbly not realistic to simulate the real world physics faithfully, in such a way that operators can use it as actual feedback. Especially in sensitive scenarios like grabbing a glass, there are tipping points (sometimes literal), where a few millimeters and 100 milliseconds is the difference between close call and full smash.
Thinking about it now, if one would deliberately add much more latency (a few seconds), it might be possible to use real-world simulation as aid. At least for operations which can be decomposed into sequences of transitions between stable/safe states. Say moving dishes from dishwasher to cupboard. Picking up is critical, but holding in hand is (presumably) safe, placing in cupboard critical, once placed it is safe. Then one could let teleoperator do the entire critical move virtually, act it out in simulator only. See what the outcome is. If high risk of failure, deny operation. If good chance of success (per simulation) can allow to execute in the real world.
More autonomous operation will need ability to simulate actions, project alternative approaches into the future, and a world model strong enough that can also plan and execute based on it. So there are potential synergies in a full-teleop, to hybrid teleop to autonomy transition.
Note, this approach would also assume the relevant environment to be static. So it would not help handle the pet or toddler...
They will brag about each unit sold to investors to demonstrate product/market fit, and then raise probably 10-100x the sales price per unit from investors.
Transformers are very powerful also for non-language data. For example time series, sequences like DNA or audio (also outside of speech and music).
Of course the vast amount of human text is key to training a typical LLM, but it is not the only use.
All language is referential. Even in everyday speech the meaning is not in the words themselves, in so much as they are pointers to concepts that (hopefully) exists already in the brains of the people we are conversing with. So when someone is very well aligned, one can convey ideas that go much further than conventionally expressed in the "general" language which is mutually intellible with most speakers of the same language. It is a rare experience though, at least for deep or personal topics.
I suspect OpenAI wants the user data from Sky/Atlas, and Apple does not want to give it. Also, Apple are probably wanting to keep their options open wrt other AI providers - after all, it is increasingly clear that OpenAI is not the only game in town for thd core LLM tech.
Lack of Accountability as-a-Service! Very attractive proposition to negligent and self-serving organizations. The people in charge don't even have to pay for it themselves, can just funnel the organization money to the vendor. Encouraging widespread adoption helps normalizes the practice. If anyone objects, shut them down as not thinking-of-the-children and something-must-be-done (and every other option is surely too complicated/expensive).
Nice. Looks like it could be a good tool for image labeling at least. But no support for audio or time-series data? :( Maybe it will get it down the line?
Makes business sense. The two companies have similar customer profiles in the embedded space, so this should make for cross selling opportunities both ways. Just being a pre-approved vendor is often a big win, as many times that involves long and costly processes.
Seems within reach of a hobbyist maker, at least the slightly unhinged kind (in the best meaning). Since considerable energy is needed but also avoiding excessive heating, integrating temperature measurement into the bulb might be prudent. Or at the very least having regulation on current+voltage. There might be risk of thermal runaway, if conductivity goes up when voltage is applied to the epoxy mixture.
I would not call validation of PCBs easy conceptually. We had to develop a lot of automation in validation to enable today production. Optical verification has been standard for many years, and X-ray is now getting commonplace also. Flying probes commonly used. Functional automated testing is standard for any non-hobby product.
But you are right in that the automation overall was/is bottlenecks on the ability to do QA - which not only prevents defects from being sent out, but is also critical to systematic improvements in the production process (both tuning and new iteration).
And I believe you are right to call out LLM based systems to be weak in this area, and it is a limiting factor. I believe that automated QA will be more and more critical to positive LLM impact.
Are you sure you spend less time per week on clothes washing than people did before the clothes washer? I suspect that before, people would wash their clothes considerably more seldom - which might even out the efficiency gains of the machine.
Their first bet was than they were going to be the frontier model provider by a good margin, and that others would not be able to compete on the "intelligence". And that they could get distribution via big customers looking to buy model access.
The dominant-model-provider strategy has already failed, many actors have models that rival them - both established (Google) and newcomers (Anthropic). Open models are not to shabby either, enough to undermine the narrative "we are uniquely able to do powerful models". As you say, there is a commodification process started, and it might be a race to the bottom in terms of margin.
So, OpenAI has moved into a new/adapted strategy, where they want to own the customers to a much larger degree, and rely less on partners/customers for distribution. This is likely because their prospective partners have a bunch of viable models to select between (many end products for power users lets people select freely), and high competitive pressure on costs (as it defines the margin and competitiveness) of the end products. Codex, Sora, their new web browser announcement, adjustments in ChatGPT is all to ensure a lot of direct end users - more brand recognition, more influence, more monetization possibilities.
So I think it is a considerable pivot from their initial plan/hopes. But it is not an unraveling - it is a rather smart response to the fierce competition in the market.
These kinds of systems tends to also be used by abusive policemen monitoring their exes, and police tracking members of undesirable (but legal) activist groups, etc.
It seems here that Google provides the core IP, and that Synaptic packages this (and probably other related IP blocks) block that can be used to build a SoC. As of now there are no chips announced. So it will be some years before we as software/electronics engineers get to play with it.
The architecture seems to be RISC-V array with standard RVV vector instruction set. That is a quite familiar environment for software developers compared to custom systolic arrays.
If done at scale, they are kinda crowd sourcing the test set from the entire internet, personal and business world. It will be harder and harder at least to pinpoint weaknesses, at least for the general public. It probably has little to do with intelligence (at least fluid intelligence as defined by Chollet et al) - but I guess it is sound tactic if the strategy is "fake it till you make it". And we might be surprised as to how far along that can go...
What development have happened in robotics over the last 2 years, that changed your mind?
I am familiar with general industry trends in electronics over the last two decades, as well as the 1 decade in machine learning.
Personally I do not think that we have seen the full battlefield impact of the last 10-20 years developments in microcontrollers, battery, power electronics, electromotors, MEMS sensors, imaging sensors, control firmware, computer vision and object detection. Capability per dollar has increased by a few orders of magnitude in each of these domains. We are seeing some hints of the potential impact now after Russia invaded Ukraine - low-cost drones are now killing and damaging just as much or more as existing high tech weaponry. These systems are crude, with limited autonomy, but it is already enough to wreck havoc. I do not think development of these systems have stopped, unfortunately.
Early day for robotics and AI still. But for a superpower, the potential for (more) autonomous weapons in that intersection is something that (unfortunately) cannot be ignored.