> Sometimes after a night’s sleep, we wake up with an insight on a topic or a solution to a problem we encountered the day before.
The current crop of models do not "sleep" in any way. The associated limitations on long term task adaptation are obvious barriers to their general utility.
> When conversing with LLMs, I never get the feeling that they have a solid grasp on the conversation. When you dig into topics, there is always a little too much vagueness, a slight but clear lack of coherence, continuity and awareness, a prevalence of cookie-cutter verbiage. It feels like a mind that isn’t fully “there” — and maybe not at all.
One of the key functions of REM sleep seems to be the ability to generalize concepts and make connections between "distant" ideas in latent space [1].
I would argue that the current crop of LLMs are overfit on recall ability, particularly on their training corpus. The inherent trade-off is that they are underfit on "conceptual" intelligence. The ability to make connections between these ideas.
As a result, you often get "thinking shaped objects", to paraphrase Janelle Shane [2]. It does feel like the primordial ooze of intelligence, but it is clear we still have several transformer-shaped breakthroughs before actual (human comparable) intelligence.
Not really, no. The founders were not omniscient, but many of them publicly wrote about the problematic rise of political "factions" contrary to the general interest: https://en.wikipedia.org/wiki/Federalist_No._10
The grid actually already has a fair number of (non-software) circular dependencies. This is why they have black start [1] procedures and run drills of those procedures. Or should, at least; there have been high profile outages recently that have exposed holes in these plans [2].
An analogy is the difference between vector and bitmap graphics.
CAD programs aren't just a different set of operations on the same data, they use an entirely different representation (b-rep [1] vs Blender's points, vertices, and polygons).
These representations are much more powerful but also much more complex to work with. You typically need a geometric kernel [2] to perform useful operations and even get renderable solids out of them.
So sure, I suppose you could build all of that into Blender. But it's the equivalent of building an entire new complex program into an existing one. It also raises major interoperation issues. These two representations do not easily convert back and forth.
So at that point, you basically have two very different programs in a trenchcoat. So far the ecosystem has evolved towards instead building two different tools that are masters of their respective domains. Perhaps because of the very different complexities inherent in each, perhaps because it makes the handover / conversion from one domain to the other explicit.
> CAD programs aren't just a different set of operations on the same data, they use an entirely different representation (b-rep [1] vs Blender's points, vertices, and polygons).
So with that in mind, there should be something that is possible to build in CAD, but impossible then to build in Blender?
I know the differences between the two, I understand they're fundamentally different, yet I seem to be able to produce similar results to others using CAD, so I'm curious what results I wouldn't be able to reproduce in Blender.
Sure. Create a diamond polygon and revolve it around a point.
Blender has methods and tools to _approximate_ doing this. It has a revolve tool... where the key parameter is the number of steps.
This is not a revolution, it's an approximation of a revolution with a bunch of planar parts.
BREP as I understand it allows you to describe the surfaces of this operation precisely and operate further on them (e.g. add a fillet to the top edge).
Ditto for things like circular holes in objects. With blender, you're fundamentally operating on a bunch of triangles. Fundamental and important solid operations must be approximated within that model.
BREP has a much richer set of primatives. This dramatically increases complexity but allows it to precisely model a much larger universe of solids.
(You can kinda rebuild functionality that geometric kernels have with geometry nodes now in blender. This is a lot of work and is not a great user interface compared to CAD programs)
This doesn't seem right to me. From the article I believe you are referencing ("What if AI made the world’s economic growth explode?"):
> If investors thought all this was likely, asset prices would already be shifting accordingly. Yet, despite the sky-high valuations of tech firms, markets are very far from pricing in explosive growth. “Markets are not forecasting it with high probability,” says Basil Halperin of Stanford, one of Mr Chow’s co-authors. A draft paper released on July 15th by Isaiah Andrews and Maryam Farboodi of mit finds that bond yields have on average declined around the release of new ai models by the likes of Openai and DeepSeek, rather than rising.
It absolutely (beyond being clearly titled "what if") presented real counterarguments to its core premise.
There are plenty of other scenarios that they have explored since then, including the totally contrary "What if the AI stock market blows up?" article.
This is pretty typical for them IME. They definitely have a bias, but they do try to explore multiple sides of the same idea in earnest.
I think any improvements to productivity AI brings will also create uncertainty and disruption to employment, and maybe the latter is greater than the former, and investors see that.
> Understanding twos complement representation is an essential programming skill
The field of programming has become so broad that I would argue the opposite. The vast majority of developers will never need to think about let alone understand twos complement as a numerical representation.
> With single modality sensors, you have no way of truly detecting failures in that modality, other than hacks like time-series normalizing (aka expected scenarios).
"A man with a watch always knows what time it is. If he gains another, he is never sure"
Most safety critical systems actually need at least three redundant sensors. Two is kinda useless: if they disagree, which is right?
EDIT:
> If multiple sensor modalities disagree, even without sensor fusion, you can at least assume something might be awry and drop into a maximum safety operation mode.
This is not always possible. You're on a two lane road. Your vision system tells you there's a pedestrian in your lane. Your LIDAR says the pedestrian is actually in the other lane. There's enough time for a lane change, but not to stop.
> Two is kinda useless: if they disagree, which is right?
They don't work by merely taking a straw poll. They effectively build the joint probability distribution, which improves accuracy with any number of sensors, including two.
> You're on a two lane road. Your vision system tells you there's a pedestrian in your lane. Your LIDAR says the pedestrian is actually in the other lane. There's enough time for a lane change, but not to stop.
Any realistic system would see them long before your eyes do. If you are so worried, override the AI in the moment.
> They don't work by merely taking a straw poll. They effectively build the joint probability distribution, which improves accuracy with any number of sensors, including two.
Lots of safety critical systems actually do operate by "voting". The space shuttle control computers are one famous example [1], but there are plenty of others in aerospace. I have personally worked on a few such systems.
It's the simplest thing that can obviously work. Simplicity is a virtue when safety is involved.
You can of course do sensor fusion and other more complicated things, but the core problem I outlined remains.
> If you are so worried, override the AI in the moment.
This is sneakily inserting a third set of sensors (your own). It can be a valid solution to the problem, but Waymo famously does not have a steering wheel you can just hop behind.
This might seem like an edge case, but edge cases matter when failure might kill somebody.
Isn't the historical voting pattern something more of a legacy thing dictated by limited edge compute of the past vs necessarily a best practice.
I see in many domains a tendency to oversimplify decision making algorithms for human understanding convenience (eg vote rather that develop a joint probability distribution in this case, supply chain and manufacturing in particular seem to love rules of thumb) rather than use better algorithms that modern compute enables higher performance, safety etc
This is an interesting question where I do not know the answer.
I will not pretend to be an expert. I would suggest that "human understanding convenience" is pretty important in safety domains. The famous Brian Kernighan quote comes to mind:
> Everyone knows that debugging is twice as hard as writing a program in the first place. So if you're as clever as you can be when you write it, how will you ever debug it?
When it comes to obscure corner cases, it seems to me that simpler is better. But Waymo does seem to have chosen a different path! They employ a lot of smart folk, and appear to be the state of the art for autonomous driving. I wouldn't bet against them.
Seatbelt mechanisms are complicated, airbag timing is complicated, let's just do away with them entirely in the name of simplicity?
No, when it comes to not killing people, I'd say that safer is usually better.
Remember the core function of the system is safety, simplicity is nice to have, but explicitly not as important.
That said, beware of calling something 'complicated' just because you don't understand it, especially if you don't have training and experience in that thing. What's more relevant is whether the people building the systems think it is too complicated.
We're trying to build vehicles that are totally autonomous, though. How do you grab the wheel of the new Waymos without steering wheels? Especially if you're in the back seat staring at Candy Crush.
Waymos are safer, and drive more defensively than humans. There is no way a Waymo is going to drive aggressively enough to get itself into the trolley problem.
This situation isn't going to happen unless the vehicle was traveling at unsafe speeds to begin with.
Cars can stop in quite a short distance. The only way this could happen is if the pedestrian was obscured behind an object until the car was dangerously close. A safe system will recognize potential hiding spots and slow down preemptively - good human drivers do this.
"Quite a short distance" is doing a lot of lifting. It's been a while since I've been to driver's school, but I remember them making a point of how long it could take to stop, and how your senses could trick you to the contrary. Especially at highway speeds.
I can personally recall a couple (fortunately low stakes) situations where I had to change lanes to avoid an obstacle that I was pretty certain I would hit if I had to stop.
At the driving school I attended, they had us accelerate to 50 mph and then slam on the brakes so we'd have a feel for the distance (and the feel).
While it's true they don't stop instantaneously at highway speeds, cars shouldn't be driving highway speeds when a pedestrian suddenly being in front of you is a realistic risk.
What if the obstacle is not a person? What if something falls off a truck in front of the vehicle? What if wildlife spontaneously decides to cross the road (a common occurrence where I live)?
I don't think these problems can just be assumed away.
You don't really ever have "two sensors" in the sense that it's two measurements. You have multiple measurements from each sensor every second. Then you accumulate that information over time to get a reliable picture. If the probability of failure on each frame were independent, it would be a relatively simple problem, but of course you're generally going to get a fairly high correlation from one frame to the next about whether or not there's a pedestrian in a certain location. The nice thing about having multiple sensing modalities is that the failure correlation between them is a lot lower.
For example, say you have a pedestrian that's partially obscured by a car or another object, and maybe they're wearing a hat or a mask or wearing a backpack or carrying a kid or something, it may look unusual enough that either the camera or the lidar isn't going to recognize it as a person reliably. However, since the camera is generally looking at color, texture, etc in 2D, and the Lidar is looking at 3D shapes, they'll tend to fail in different situations. If the car thinks there's a substantial probability of a human in the driving path, it's going to swerve or hit the brakes.
> > If multiple sensor modalities disagree, even without sensor fusion, you can at least assume something might be awry and drop into a maximum safety operation mode.
> This is not always possible. You're on a two lane road. Your vision system tells you there's a pedestrian in your lane. Your LIDAR says the pedestrian is actually in the other lane. There's enough time for a lane change, but not to stop.
> What do you do?
Go into your failure mode. At least you have a check to indicate a possible issue with 2 signals.
I came here to write the same comment you did. What I’d suspect (I don’t work in self driving but I do in AI) is the issue is that this mode of operation would happen more often than not as the sensors disagree in critical ways more often than you’d think. So going “safety first” every time likely critically diminishes UX.
The issue is not recognising that optimising for Ux at the expense of safety here is the wrong call, motivated likely by optimism and a desire for autonomous cars, more than reasonable system design. I.e. if the sensors disagree so often that it makes the system unusable, maybe the solution is “we’re not ready for this kind of technology and we should slow down” rather than “let’s figure out non-UX breaking edge case heuristics to maintain the illusion of autonomous driving being behind the corner”.
Part of this problem is not even technological - human drivers tradeoff safety for UX all the time - so the expectation for self driving is unrealistic and your system has to have the ethically unacceptable system configuration in order to have any chance of competing.
Which is why - in my mind - it’s a fools endeavour in personal car space, but not in public transport space. So go waymo, boo tesla.
Exactly my point. That you know the systems disagree is a benefit, compared to a single system.
People are underweighting the alternative single system hypothetical -- what does a Tesla do when its vision-only system erroneously thinks a pedestrian is one lane over?
> This is not always possible. You're on a two lane road. Your vision system tells you there's a pedestrian in your lane. Your LIDAR says the pedestrian is actually in the other lane. There's enough time for a lane change, but not to stop.
This is why good redundant systems have at least 3... in your scenario, without a tie-breaker, all you can do is guess at random which one to trust.
That's a good point, but people do need to keep in mind that many engineered systems with three points of reference have three identical points of reference. That's why it works so well, a common frame of reference (i.e. you can compare via simple voting).
For example jet aircraft commonly have three pitot static tubes, and you can just compare/contrast the data to look for the outlier. It works, and it works well.
If you tried to do that with e.g. LIDAR, vision, and radar with no common point of reference, solving for trust/resolving disagreements is an incredibly difficult technical challenge. Other variations (e.g. two vision + one LIDAR), does not really make it much easier either.
Tie-breaking during sensor fusion is a billion+ dollar problem, and will always be.
The reality is also that nobody (aside from Mark "I Want To Buy a State of the Art AI Research Lab" Zuckerberg) is even offering millions in cold hard cash.
Instead, they're offering something worse: the _chance_ to cash out equity that _might_ be worth that at _some_ point in the future.
Versus spending time with my kid right now. Or any of the hundreds of other more enjoyable things I can do with my time.
They're dangling a lottery ticket in front of us. I've seen the end of that movie several times myself now; enough to know the odds are long.
The advice I've seen with delegation is the exact opposite. Specifically: you can't delegate what you can't do.
Partially because of all else fails, you'll need to step in and do the thing. Partially because if you can't do it, you can't evaluate whether it's being done properly.
That's not to say you need to be _as good_ at the task as the delegee, but you need to be competent.
For example, this HBR article [1]. Pervasive in all advice about delegation is the assumption that you can do the task being delegated, but that you shouldn't.
> Just that it's not an expectation, e.g., you don't expect a CEO to be able to do the CTO's job.
I think the CEO role is actually the outlier here.
I can only speak to engineering, but my understanding has always been that VPs need to be able to manage individual teams, and engineering managers need to be somewhat competent if there's some dev work that needs to be done.
This only happens as necessary, and it obviously should be rare. But you get in trouble real quickly if you try to delegate things you cannot accomplish yourself.
The current crop of models do not "sleep" in any way. The associated limitations on long term task adaptation are obvious barriers to their general utility.
> When conversing with LLMs, I never get the feeling that they have a solid grasp on the conversation. When you dig into topics, there is always a little too much vagueness, a slight but clear lack of coherence, continuity and awareness, a prevalence of cookie-cutter verbiage. It feels like a mind that isn’t fully “there” — and maybe not at all.
One of the key functions of REM sleep seems to be the ability to generalize concepts and make connections between "distant" ideas in latent space [1].
I would argue that the current crop of LLMs are overfit on recall ability, particularly on their training corpus. The inherent trade-off is that they are underfit on "conceptual" intelligence. The ability to make connections between these ideas.
As a result, you often get "thinking shaped objects", to paraphrase Janelle Shane [2]. It does feel like the primordial ooze of intelligence, but it is clear we still have several transformer-shaped breakthroughs before actual (human comparable) intelligence.
1. https://en.wikipedia.org/wiki/Why_We_Sleep 2. https://www.aiweirdness.com/