I also don't believe in the 'bitter lesson' when extrapolated to apply to all 'AI application layer implementations' - at least in the context of asserting that the universe of problem scopes are affected by it.
I think it is true in an AI research context, but an unstated assumption is that you have complete data, E2E training, and the particular evaluated solution is not real-world unbounded.
It assumes infinite data, and it assumes the ability to falsify the resulting model output. Most valuable, 'real world' applications of AI when trying to implement in practice have an issue with one or both of those. So in other words: where a fully unsupervised AI pathway is viable due to the structure of the problem, absolutely.
I'm not convinced in the universality of this. Doesn't mean the core point of this essay on the futility of startups basing their business around one of the off the shelf LLMs isn't valid - I think for many they risk being generalized away.
Li-ion is a very big umbrella of battery chemistries, and within can have broadly different safety and stability profiles.
More important is that Li-ion cells are no joke, and are not equivalent to ‘AA’ or ‘AAA’ battery formats. You can NCM, NCA, LFP, Cobalt chemistries and blends within - not even accounting for differing mAH and charge/discharge capacities. 18650 is simply a form factor.
Better labeling, specing, regulated and certified repair processes are just a start for solutions as it is currently a Wild West.
However, statements like ‘Li-ion should be replaced ASAP’ do not factor in that… lithium ion as a platform is not going anywhere and has essentially won at least the next half century for 18650 type form factors and applications. Moving away is not a reasonable solution.
You are missing the core technology point that drives the underlying thesis here - ‘every manufactured good, gets cheaper per unit as you make more and more of it’.
Whether this is from the lab, a recipe of plant inputs, or some other currently unknown thing - if you are able to develop and commercialize a product around a ‘specific technology’, you can drive the cost of that down through experience and scale.
Animals are a VERY thermodynamically inefficient way of converting plants to ‘high quality proteins’. We have automated the crap out of the food processing system. There are no material efficiencies left to be had. Compare that to options which are on paper thermodynamically superior (I.e. not supporting the life of an animal to only use their muscle tissue).
By taking a technology with a much higher theoretical efficiency, and then scaling that up - creates the classic technology disruption scenario. ->‘It’s the cheaper version of X commodity, why wouldn’t I choose that?’
That is because CdTe is actually a stable molecule that isn’t water soluble, turns out the problem of panels from a toxic leeching perspective is largely from the small amounts of lead used at soldering points which is a common point with most silicon PV panels.
No reason tin can’t be used instead/lead free panels can’t be made except for the saving of a few pennies by skimping on the solder - no regulations to ensure they should be lead free. Unfortunately waste processing decades into the future isn’t often accounted for and those few pennies at design time add up across hundreds of thousands of panels on a single site…
Not to be incredibly nit-picky, but the costs of the product right now don’t matter nearly as much as the ability to scale production and variable unit costs.
As a ‘hard tech’, they need to work themselves through the technology readiness framework before the cost of a particular material even tells you anything of value.
‘Sustainable’ in the food context is much more around total resource use (land, water, materials), waste products (ghgs and waste water) rather than a primary focus of energy efficiency.
Besides - heatpumps are largely the exception in thermodynamics, expecting anything more than 100% efficiency for most non-heat moving applications should not be expected.
Capturing CO2 is not ‘energy cheap’ and therefore it is best when and where possible much of what we start capturing to rid it from the carbon cycle rather than re-introduce it.
a) it will not necessarily last longer, li-ion in the right conditions (particularly more stable chemistries) have superior cycle life experience in practice
b) yes, newest batteries decay slowly over time and don’t ‘just die’ - but this is the same as li-ion
c) safety is mitigated in the modern setting by utilizing the same charge controllers used to keep EV car batteries safe. Remember most North American homes with a natural gas furnace literally is a controlled blast of explosive gas without a smell that people just have and don’t think about - as the tech goes on the learning curve the management issues (I.e. disconnecting faulty cels individually) seems like a plausible end solution, even if that is necessary.
Lastly, while NiMH does not use soon to be very scare processed lithium, NiMH uses very large amounts of nickel, cobalt and comparatively exotic ‘rare earth metals’ who’s production should be expected to be even more challenging to scale.
All that said - I am all for us scaling up all of these technologies so that each can fit within a particular market niche based on inherent pros and cons.
Hypothetically this is not a binary option, they could significantly ramp up the strictness of 3rd party sellers increasing the costs of making multiple, temporary fake seller accounts more challenging.
That said though - this would clearly impact profits proportional to the level of strictness/barriers.
For getting an intuition on it, I would recommend the clip from HBO’s Chornobyl which ELI5’s a nuclear meltdown. [1]
If the reactor is air cooled, that can be translated as “the heat can just burn off” as there isn’t a liquid cooling system, reducing this failure mode.
As mentioned in the video, reactions are about balance. If a ‘safe’ SMR gets shut off or has its systems fail, you would want the accelerators / enablers of the reaction to disappear, and roll out the reaction.
This is unlike most fission reactors in use today which are a ‘balanced dance’ where lots of the design is focused on reducing the speed of the burn (like a constant stream of water on a fire) rather than say controlling the oxygen of a fire with limited fuel. SMRs aim be more of the “control the oxygen” type of reaction rather than “keep the fire cool” kind of reaction. (Forgive my vast oversimplification, but this is my intuition on it).
As someone who never got “hooked” onto twitter (turned into an active user) each time I viewed Twitter I found two main issues:
- a discoverability problem for topics, authors, and tweets
- too much garbage, spam, low quality tweets
To the point where I “churned out”.
It is clear that twitter as a platform has immense long term potential if curated properly. The fact that “cancel tribes” and virtual lynchings are a recent mainstay of the culture of the user base, shouldn’t make it surprising that critical and interesting voices (no free speech) do not feel free. Enabled wokism from the top down has materially affected the quality of the content on the platform in its current form.
The issue at twitter is likely a combination of:
- poor management
- internal cultural issues
- lack of a revamped product vision
All of the above issues are the perfect set up for an executive shakeup from an outsider.
If we take the above as true, who else has the gull and ability to do such a shake-up? Twitter’s board as demonstrated in the previous weeks seemed quite entrenched and reasonably powerful.
TCOs for individual consumers will be vastly different depending on the parameters of the situation, with particular impact driven by volume of miles driven annually.
Due to this, commercial fleets in my view have a much clearer picture here than personal vehicle decisions. This is also particularly driven by personal vehicle purchases being driven less by “TCO” in a strict sense but rather “what is my monthly payment”.
If I may riff and provide a potential view of a tangential alternative force that was at play that made the project destined to be doomed (via this approach you mention about un-implemented features across lower cost legacy hardware)
The declining costs of the hardware (screens, capacitive touch, battery, processor, memory, etc etc) via a Wright's Law/ Moore's Law type curve meant that over-optimizing around lowering the costs by choosing 'imperfect' hardware meant it was very difficult to ever get this going as a viable platform - which if we imagine a world of these costs declining 1/4 the rate (4x slower), this ecosystem would have more likely found enough backers and funding to get some more roots and sprouts going.
Ultimately, the cost curve declines of the other equipment meant that smartphones/Netbooks/tablets were bound to leapfrog it well before this OLPC 'formulation' was able to take off and be sustainable on its own.
Can't say that 'it is the end' yet - but given that pandemics end when there is no more 'room' for the virus to move through naive hosts (places to infect). The fact that Omicron is SO contagious means it will effectively find any pockets of 'fuel' in the adult human population in the coming weeks and months. Everyone will be exposed.
It is probably more accurate to say it is the likely start of the transition between 'pandemic' and 'endemic' phase. We say endemic as it LOOKS like the gap between coronavirus immunity length to viral instability (number of changes) is longer than for influenza. (T-cell response still very strong, original SARS patients almost 2 decades later have cross immunity to SARS-Cov-2, Omicron immunity effective against delta, etc) That said - it is VERY unlikely to be eradicable at this point due to its ability to infect so many different species and circulate and breed in the wild (more so than even influenza).
So new variants down the line in the coming years/decades is more likely (not different than swine flu, bird flu, etc pandemic risks in the recent past). But, if longer term immunity holds up - none of these will have anywhere near the impact given the absence of billions of naive hosts to burn through.
So if we call the transition from pandemic to endemic the end (likely, yes) - then probably.
An attempt at a serious answer now that it looks like you fixed the units in your post.
One key part is recognizing that “energy TRANSITION” accounts inherently that we are not talking about the design of the final end state 50-100 years from now, as doing so is not useful to the conversation as there are too many unknowns that far ahead. Better to recognize that you do not need 1:1 mapping of energy storage to power usage requirements in this “toy” example as fulfilled by li-ion production.
In practice:
- Solar and wind generation compliment each other structurally (sunnier when less windy, windier at night, seasonally too) meaning that in practice you only need a fraction of power capacity per amount of solar and wind generation. The best summary I’ve come across for this is some of Tony Seba’s/RethinkX’s work on “super power”[1]
- it is always sunny or windy somewhere. More renewables increase the incentive for deeper and longer distance grid interconnections, meaning continental differences in generation can be “smoothed out”
- li-ion energy storage is great for same day fluctuations(storage measured in hours), less so for longer. Other energy storage tech platforms have different economic properties leading to more sensible deployments (the other end of the spectrum is making hydrogen gas via electrolyzer to then convert back via fuel cell - best for storage measured in months)
- You also need to factor in the nuclear and hydroelectric sources existing 50 years from now, as well as the natural gas to bridge the variability until then.
Could go on but there is lots out there discussing the involvement and feasibility of having a very renewables heavy grid.
The key thing is that we need to build as much of all renewables, in as many places as possible ASAP. Penetration of both solar and wind as we can already see can operate well on grids with virtually 0 energy storage. Worst case you can always curtail production. Curtailed production represents someone’s energy storage opportunity.
I think thread parents point is that the problems are obtuse and plentiful enough that in order to understand them you need to be involved in point 1, which is arguably the hard part.
When thinking about “carbon stores” - it is significantly more preferable to turn managed forests into manufactured wood products for buildings and furniture (where the CO2 gets functionally stored and then possibly tossed in a landfill far in the future) to turning those forests into wood pellets to be burned as biomass.
Managed forestry has the ability to, relatively trivially and particularly when compared to other “carbon sequestration alternatives” lock in decent magnitudes of carbon as we figure out our energy system.
Particularly helps when comparing to other building material options (bricks, concrete, steel all very energy intensive and not decarbonized yet).
Abundant renewable energy, driven by decreases in costs of solar, wind and energy storage could be used with no other inputs but energy* via the Haber process + electrolysis.
I suppose I don’t know the chemical compound that constitute “solid soil amendment” though. If it includes functionally free nitrogen/ammonia and proper management of land or equipment, is this still not possible?
I can't help but think that this action could very well spur out some 'unintended consequences'.
Prohibition can restrict and reduce participation and usage at an overall level, but certainly some portion of this will be pushed underground, or to unregulated spaces - probably just the thing the government WOULDN'T want to have.
Particularly when we are talking about some portion of the players being addicts here... how many are going to sign up for VPNs, Tor (or some other workaround) and find some other game or activity that satisfies that itch?
I think it is true in an AI research context, but an unstated assumption is that you have complete data, E2E training, and the particular evaluated solution is not real-world unbounded.
It assumes infinite data, and it assumes the ability to falsify the resulting model output. Most valuable, 'real world' applications of AI when trying to implement in practice have an issue with one or both of those. So in other words: where a fully unsupervised AI pathway is viable due to the structure of the problem, absolutely.
I'm not convinced in the universality of this. Doesn't mean the core point of this essay on the futility of startups basing their business around one of the off the shelf LLMs isn't valid - I think for many they risk being generalized away.