Foyle

Karma: 149

Foyle Sep 11, 2024, 8:55 AM
3 points
0
in reply to: Daniel Kokotajlo’s comment on: Self-driving car bets
They cannot just add an OOM of parameters, much less three.
How about 2 OOM’s?
HW2.5 21Tflops HW3 72x2 = 72 Tflops (redundant), HW4 3x72=216Tflops (not sure about redundancy) and Elon said in June that next gen AI5 chip for fsd would be about 10x faster say ~2Pflops
By rough approximation to brain processing power you get about 0.1Pflop per gram of brain so HW2.5 might have been a 0.2g baby mouse brain, HW3 a 1g baby rat brain HW4 perhaps adult rat, and upcoming HW5 a 20g small cat brain.
As a real world analogue cat to dog (25-100g brain) seems to me the minimum necessary range of complexity based on behavioral capabilities to do a decent job of driving—need some ability to anticipate and predict motivations and behavior of other road users and something beyond dumb reactive handling (ie somewhat predictive) to understand anomalous objects that exist on and around roads.
Nvidia Blackwell B200 can do up to about 10pflops of FP8, which is getting into large dog/wolf brain processing range, and wouldn’t be unreasonable to package in a self driving car once down closer to manufacturing cost in a few years at around 1kW peak power consumption.
I don’t think the rat brain HW4 is going to cut it, and I suspect that internal to Tesla they know it too, but it’s going to be crazy expensive to own up to it, better to keep kicking the can down the road with promises until they can deliver the real thing. AI5 might just do it, but wouldn’t be surprising to need a further oom to Nvidia Blackwell equivalent and maybe $10k extra cost to get there.

Foyle Jun 28, 2024, 4:27 AM
2 points
0
on: How Big a Deal are MatMul-Free Transformers?
There has been a lot of interest in this going back to at least early this year and the 1.58bit LLM (ternary) logic paper https://arxiv.org/abs/2402.17764 so expect there has been a research gold rush and a lot of design effort going into producing custom hardware almost immediately that was revealed.
With Nvidia dual chip GB200 Grace Blackwell offering (sparse) 40Pflop fp4 at ~1kW there has already been something close to optimal hardware available—that fp4 performance may have been the reason the latest generation Nvidia GPU are in such high demand—previous generations haven’t offered it as far as I am aware. For comparison a human brain is likely equivalent to 10-100Pflops, though estimates vary.
Being able to up the performance significantly from a single AI chip has huge system cost benefits.
All suggesting that the costs for AI are going to drop yet again, and human level AGI operating costs are going to be measured in cents per hour when it arrives in a few years time.
The implications for autonomous robotics are likely tremendous, with potential OOM power savings likely to bring far more capable systems to smaller platforms, home robotics, fsd cars, and (scarily) military murderbots. Tesla has (according to Elon comments) a new HW5 autonomy chip coming out next year that is ~50x faster than their current FSD development baseline HW3 2 x 72Tflop chipset, but needs closer to 1kW power, so they will be extremely keen on implementing something that could save so much power.

Foyle Jun 23, 2024, 11:14 AM
0 points
1
in reply to: Stephen Fowler’s comment on: Nathan Young’s Shortform
AI safety desperately needs to buy in or persuade some high profile talent to raise public awareness. Business as usual approach of last decade is clearly not working—we are sleep walking towards the cliff. Given how timelines are collapsing the problem to be solved has morphed from being a technical one to a pressing social one—we have to get enough people clamouring for a halt that politicians will start to prioritise appeasing them ahead of their big tech donors.
It probably wouldn’t be expensive to rent a few high profile influencers with major reach amongst impressionable youth. A demographic that is easily convinced to buy into and campaign against end of the world causes.

Foyle May 16, 2024, 2:05 AM
1 point
0
in reply to: Steve Kommrusch’s comment on: Super-Exponential versus Exponential Growth in Compute Price-Performance
Current Nvidia GPU prices are highly distorted by scarcity, with profit margins that are reportedly in the 80-90% of sale price range: https://www.tomshardware.com/news/nvidia-makes-1000-profit-on-h100-gpus-report
If these were commodified to the point that scarcity didn’t influence price then that $/flop point would seemingly leap up by an order of magnitude to above 1e15Flop/$1000 scraping the top of that curve, ie near brain equivalence computation power in $3.5k manufactured hardware cost, and latest Blackwell GPU has lifted that performance by another 2.5x with little extra manufacturing cost. Humans as useful economic contributors are so screwed, even with successful alignment the socioeconomic implications are beyond cataclysmic.

Foyle Feb 18, 2024, 2:03 AM
2 points
0
on: Raising children on the eve of AI
I’m going through this too with my kids. I don’t think there is anything I can do educationally to better ensure they thrive as adults other than making sure I teach them practical/physical build and repair skills (likely to be the area where humans with a combination of brains and dexterity retain useful value longer than any other).
Outside of that the other thing I can do is try to ensure that they have social status and financial/asset nest egg from me, because there is a good chance that the egalitarian ability to lift oneself through effort is going to largely evaporate as human labour becomes less and less valuable, and I can’t help but wonder how we are going to decide who gets the nice beach-house. If humans are still in control of an increasingly non-egalitarian world then society will almost certainly slide towards it’s corrupt old aristocratic/rentier ways and it becomes all about being part of the Nomenklature (communist elites).

Foyle Feb 9, 2024, 7:01 AM
13 points
0
on: Running the Numbers on a Heat Pump
[disclaimer: I am a heat pump technology developer, however the following is just low-effort notes and mental calcs of low reliability, they may be of interest to some. YMMV]
It may be better to invest in improved insulation.
As rough rule of thumb COP is = eff * Theat/(Theat-Tcold), with Temperatures measured in absolute degrees (Kelvin or Rankine), eff for most domestic heat pumps is in range 0.35 to 0.45, high quality european units are often best for COP due to long history of higher power costs—but they are very expensive, frequently $10-20k
Looking at the COP for the unit you quoted the eff is only about 0.25 at rated conditions, not good, unless you get a much larger unit and run it at a less powerful more efficient load point.
That’s a pretty huge electricity price, about 4.5x gas price (which is distorted-market nuts, 3x is more usual globally). Given that differential it might be better to look at an absorption heat pump like https://www.robur.com/products/k18-simplygas-heat-pump that gives up to 1.7x gas heat—though they look to be on the order of $10k.
Here’s an annoying fact; If you ran that $2/therm gas (~$0.07/kWh) through a reasonably efficient (~40%) natural gas genset it would produce electricity cheaper than what you currently pay for power, and you would have 2/3rds of the gas energy left over as heat. A genset in your neighbourhood could provide a few 10′s of houses with cheaper electricity and low cost waste heat, though no doubt prevented by regulatory issues. There are a few small combined heat and power (CHP) domestic units on the market, but they tend to be very expensive, more tech-curios than economically sensible.

Foyle Jan 31, 2024, 9:42 PM
1 point
−2
on: on neodymium magnets
Niron’s Fe16N2 looks to have a maximum energy product (figure of merit for magnet ‘strength’ up to 120 MGOe at microscopic scale, which is about double that of Neodymium magnets (~60), however only 20 MGOe has been achieved in fabrication. https://www.sciencedirect.com/science/article/am/pii/S0304885319325454
Processing at 1GPa and 200°C isn’t that difficult if there is commercial benefit. Synthetic diamonds are made in special pressure vessels at 5GPa and 1500°C. There is some chance that someone will figure out a processing route that makes it possible to achieve bulk crystal orientations that unlocks higher energy products—potential payoff is huge. I expect AGI and ASI will figure out a lot of major improvements in materials science over next 1-2 decades.

Foyle Dec 13, 2023, 12:44 AM
5 points
0
on: Significantly Enhancing Adult Intelligence With Gene Editing May Be Possible
I read of a proposal a few months back to achieve brain immortality via introduction of new brain tissue that can be done in a way as to maintain continuity of experience and personality over time. Replenisens , Discussion on a system for doing it in human brains That would perhaps provide a more reliable vector for introduction, as the brain is progressively hybridised with more optimal neural genetic design. Perhaps this could be done more subtly via introduction of ‘perfected’ stem cells and then some way of increasing rate of die off of old cells.
Instead of gene editing could you just construct a ‘perfect’ new chromosome and introduce one or more instances of it into existing neurons via viral injection techniques to increase expression of beneficial factors? No particular reason why we can only have 42 🤣46 chromosomes, and this would perhaps side-step difficulties to do with gene editing. Might be a more universal solution too if we could come up with a single or small variety of options for a ‘super’ brain optimising added chromosome.
Politically the way to pitch it would be for its life saving/enhancement ability—offered for example to people with low intelligence and educational outcomes to offer them a better chance at happiness in life.

Foyle Nov 25, 2023, 11:18 AM
3 points
0
in reply to: ChristianKl’s comment on: Why not electric trains and excavators?
“So your job depends on believing the projections about how H2 costs will come down?”
I wouldn’t waste my life on something I didn’t see as likely—I have no shortage of opportunities in a wide variety of greentech fields. Hydrogen is the most efficient fuel storage ‘battery’ with 40-50% round-trip energy storage possible. Other synthetic fuels are less efficient but may be necessary for longer term storage or smaller applications. For shipping and aviation however LH2 is the clear and obvious winner.
Desert pv will likely come down in price to consistent ~$0.01-0.02 in next decade with impact of AI on manufacturing, installation and maintenance costs (a few large pv installations are already contracted in this cost range). And electrolysis and liquefaction tech are on track to yield the stated $1.50/kg (learning curves are magic). That ‘stranded’ desert pv power needs to be delivered to far distant users and hydrogen pipelines or shipping provides most realistic option for doing that.
Capturing carbon for synthetic hydrocarbons is not a trivial issue/cost. And their round trip energy storage efficiencies for synthetics hydrocarbons are worse than for hydrogen. There will still be some applications where they make the most sense. Ammonia might work too, though it also needs hydrogen feedstock and is often lethal when inhaled.
But in general I see a pretty clear path to renewable hydrogen undercutting fossil fuels on cost in the next decade or two, and from there a likely rapid decline in their use - so reasons for optimism about energy part of our civilisational tech stack at least, without breakthroughs in nuclear being needed.

Foyle Nov 24, 2023, 12:27 AM
3 points
0
in reply to: bhauth’s comment on: Why not electric trains and excavators?
Battery augmented trains: Given normal EV use examples, Tesla et al, and Tesla Semi a charging time of 10% of usage time is relatively normal. Eg charging for 20 minutes and discharging for 3 hours, or in Tesla Semi’s case might be more like an hour for 8-10 hours operation but trains have lower drag (and less penalty for weight) than cars or trucks so will go further for same amount of energy. The idea is therefore that you employ a pantograph multi MW charging system on the train that only needs to operate about 10% of the time,. This may reduce electrification capital and maintenance costs. Another option would be battery engines that can join up or disconnect from trains in motion between charging cycles in sidings.
CATL (largest battery producer in the world) are innovating heavily on Sodium ion batteries and have stated that they believe cost per kWh can come down to $40/kWh as they scale up manufacture on their second generation. “CATL first-generation sodium-ion cells cost about 77 USD per kWh, and the second generation with volume production can drop to 40 USD per kWh.”
I work professionally developing Liquid hydrogen fueled transport power technology. It’s very practical for some applications, particularly aircraft and ships that I expect will transition to hydrogen in next 2-3 decades, and possibly trains and agriculture. Using low cost power sources such as grid scale pv in low cost desert regions the price is expected to come down over next 1-2 decades to being competitive or even undercutting fossil fuels (commonly expected/roadmapped to be ~$1.5/kg H2, which translates to about $0.10/kWh useful output power, similar to fossil fuels before taxes). This is likely the only realistic route to fully renewable power for human civilisation—producing in cheapest sunny or windy areas and using at high latitudes/through winters. But LH2 is difficult to transport, transfer and employ at small scales due to lack of economic cryocooling and complexity, insulation and cost scale (r² vs r³) issues with tanks (LH2 and GH2 are very low density) and GH2 is even worse for non-pipeline distribution, so a more convenient dense and long-term easily and cheaply stored energy carrier such as Ammonia or synthetic hydrocarbons made using future cheap hydrogen feedstocks may be a better option. That is especially true for off-road uses.
Agriculture is particularly intractable, cropping farmers in my region frequently run 100+ hours a week using many MW of power during harvest which is a terrible issue for power economics without exceptionally cheap energy storage—liquid fuel is probably the only economic option.
And yes I am well aware of mining electrification, but I am very suspicious of it’s utility given many cases will see it powered by fossil fuel generation. Likely that a lot of it is PR greenwashing rather than effective CO2 reduction.

Foyle Nov 21, 2023, 11:05 PM
1 point
−2
on: Why not electric trains and excavators?
Battery electric trains with a small proportion of electrified (for charging) sections seems like a decent and perhaps more economic middle ground. Could get away with <10% of rail length electrified, and sodium batteries are expected to come down to ~$40/kWh in next few years. High utilisation batteries that are cycled daily or multiple times a day have lower capital costs. May also work for interstate trucking.
Earth moving electrification is probably the last application that makes sense or needs focusing upon, due to high capital costs of electrification and low equipment utilisation (a lot of it spends only a few percent of year being used), as well as difficulty in getting electric power to them in difficult to access off-grid locations.
Farm equipment is more important, but incredibly bad economics due to high peaks (using up to several MW to run a few large machines for a few days several times a year for cropping) but very low average utilisation.
Both of these are probably best served long term by some renewable liquid fuel and IC engines, for high power, low mass-fuelling and low capital costs. Synthetic hydrocarbon, Ammonia or Liquid Hydrogen.

Foyle Nov 5, 2023, 6:14 AM
2 points
0
on: If AGI is imminent, why can’t I hail a robotaxi?
Insufficient onboard processing power? Tesla’s HW3 computer is about 70Tflops, ~0.1% of estimated 100Pops human brain. Approximately equivalent to a mouse brain. Social and predator mammals that have to model and predict conspecific and prey behaviors have brains that generally start at about 2% human for cats and up to 10% for wolves.
I posit that driving adequately requires modelling, interpreting and anticipating other road users behaviors to deal with a substantial number of problematic and dangerous situations; like noticing erratic/non-rule compliant behavior in pedestrians and other road users and adjusting response, or negotiating rule-breaking at intersections or obstructions with other road users, and that is orders of magnitude harder than strict physics and road rule adherence. This is the inevitable edge-case issue that there are simply too many odd possibilities to train for and greater innate comprehension of human behaviors is needed.
Perhaps phone-a-smarter-AI-friend escalation of the outside-of-context problems encountered can deal with much of this, but I think cars need a lot more on board smarts to do a satisfactory job. A cat to small-dog level H100 @4Pflops FP8 might eventually be up to the task.
Elon’s recent announcements about end-to-end neural nets processing for V12.xx FSD release being the savior and solution appears to not have panned out—there’s been months of silence after initial announcement. That may have been their last toss of the die for HW3 hardware. ~~Their next iteration HW4 chip being fitted to new cars is apparently ~8x faster at 600Tflops~~.[edit, the source I found appears unreliable, specs not yet public] They really wouldn’t want to do that if they didn’t think it was necessary.

Foyle Oct 26, 2023, 7:36 AM
−1 points
4
on: Architects of Our Own Demise: We Should Stop Developing AI
Global compliance is the sine qua non of regulatory approaches, and there is no evidence of the political will to make that happen being within our possible futures unless some catastrophic but survivable casus belli happens to wake the population up—as with Frank Herbert’s Butlerian Jihad (irrelevant aside; Samuel Butler who wrote of the dangers of machine evolution and supremacy lived at film location for Eddoras in Lord of the Rings films in the 19th century).
Is it insane to think that a limited nuclear conflict (as seems to be an increasingly likely possibility at the moment) might actual raise humanities chances of long term survival—if it disrupted global economies severely for few decades and in particular messed up chip production.
What links here?
- jdp's comment on AI as a science, and three obstacles to alignment strategies by So8res (Oct 26, 2023, 9:23 AM; 112 points)

Foyle Oct 26, 2023, 7:14 AM
4 points
−1
on: Thoughts on responsible scaling policies and regulation
Any attempts at regulation are clearly pointless without global policing. And China, (as well as lesser threats of Russia, Iran and perhaps North Korea) are not going to comply no matter what they might say to your face if you try to impose it. These same issues were evident during attempts to police nuclear proliferation and arms reduction treaties during the cold war, even when both sides saw benefit in it. For AI they’ll continue development in hidden or even mobile facilities.
It would require a convincing threat of nuclear escalation or possibly a rock-solid economic blockade/embargo of non-transparent regimes to make the necessary all-the-time-and-everywhere access compliance policing happen. The political will for these options is nil. No Politician wants to escalate tensions, they are highly risk averse, and they are not going to be able to sell that to the public in a democracy. Our democracies do not select for leaders that would be capable of holding such a line.
So without instituting liberal democracies everywhere with leadership that genuinely puts humanities future ahead of their national interests this line of seeking to slow development to research alignment via regulation seems rather pointless.
Humanity needs a colossal but survivable AI scare ASAP if it is to develop the collective will to effectively regulate AI development, and not sleep walk its way off the cliff edge of extinction as seems to be our current lazy and disinterested path.

Foyle Sep 12, 2023, 10:49 PM
1 point
0
on: Is there something fundamentally wrong with the Universe?
Nothing wrong with the universe, from an Anthropic perspective it’s pretty optimal, we just have most humans running around with much of their psychology evolved to maximize fitness in highly competitive resource limited hunter-gatherer environments, including a strong streak of motivating unhappiness with regard to things like; social position, feelings of loneliness, adequacy of resources, unavailability of preferred sex partners, chattel ownership/control, relationships etc and a desire to beat and subjugate most dangerous competitors to get more for ourselves (men wanting to take down the men in the tribe next door who are likewise planning to murder them, women wanting to take down women outside of their immediate friend/family circle that compete for essential resources they need for their kids). We are designed to be to some degree miserable and discontent, with maladapted compulsions to make us work harder on things that don’t have immediate payoff, but that do improve our chances of survival and successful procreation over long term.
The fix would be hacking human brains (benignly) and figuring out how to rewire the innate negative psychological drives to enable greater contentment in a post-scarcity technological world. There’s a good chance that will become possible post singularity (if humans aren’t all dead)

Foyle Aug 19, 2023, 3:41 AM
2 points
0
on: That time I went a little bit insane
Interesting.
As counterpoint from a 50 year old who has struggled with meaning and direction and dissatisfaction with outcomes (top 0.1% ability without as yet personally satisfactory results) I have vague recollections of my head-strong teen years when I felt my intellect was indomitable and I could master anything and determine my destiny though force of will. But I’ve slowly come to the conclusion that we have a lot less free-will than we like to believe, and most of our trajectory and outcomes in life are set by neurological determinants—instincts and traits that we have little to no control over. Few if any have drive necessary (another innate attribute) to overcome or modify their base tendencies over long term, a sisyphean endeavour at best, and the likely reason we see such heavy use of nootropic drugs.
Without an (almost literal) gun to your head you can only maintain behavioural settings at odds with your likely genetically innate preferences for short periods, months, perhaps years, but not decades, ‘fake it till you make it’ in personality modification mostly doesn’t work—in my experience people (myself included) almost always fall back into their factory settings. Choose your goals, careers and partners accordingly, follow your interests, and try to find work environments where others strengths can compensate for deficits you perceive in yourself, eg working with some highly conscientious types is a massive boon if you are ADHD and probably leads to greater satisfaction and better results.

Foyle Aug 19, 2023, 12:58 AM
10 points
0
on: The U.S. is becoming less stable
I don’t think it’s mild. I’m not American, but follow US politics with interest.
A majority of blue-collar/conservative US now see the govt as implacably biased against their interests and communities, eg see recent polling on attitudes towards DOJ, FBI eg https://harvardharrispoll.com/wp-content/uploads/2023/05/HHP_May2023_KeyResults.pdf
There is widespread perception that rule of law has failed at the top levels at least—with politically motivated prosecutions (timing of stacked Trump indictments is clearly motivated by his candidacy) and favored treatment of Hunter (majority of US population see corruption in Biden’s VP doings).
And worst of all a substantial fraction of US do not believe results of 2020 elections were legitimate. In recent polling: “61% of Americans say Biden did legitimately win enough votes to win the presidency, and 38% believe that he did not”.
The traditionally middle seeking and sense-making Media has become implacably polarised, with steadfast silence on or spinning of stories inconvenient to their ‘side’.
The greatest concern must be that the rate of descent into extreme political and institutional Tribalism and differential treatment on that basis. Long term that can only lead to a violent collapse in rule of law, or repressive policing.
The only antidote is dogmatic enforcement of even-handed treatment at all levels, govt employees must be seen to be scrupulously non-partisan or everything will inevitably fall apart.

Foyle Aug 9, 2023, 3:54 AM
11 points
10
on: Perpetually Declining Population?
Ruminations on this topic are fairly pointless because so many of the underpinning drivers are clearly subject to imminent enormous change. Within 1-2 generations technology like life extension, fertility extension, artificial uteruses, superintelligent AI, AI teachers and nannies, trans-humanism and will render meaningless today’s concerns that currently seem dominating and important (if Humans survive that long). Existential risk and impacts of AI are really the only issues that matter. Though I am starting to think that the likely inevitable next generation of MAD—space based nukes hanging like an undetectable sword of Damocles over our heads is scary as hell too—coordinated global 1st strikes with only 2-3 seconds between detectability and detonation of stealth bombs re-entering from far distant orbits at >10km/s.
But even in their absence if there was close to a technology level freeze starting now evolution would move in to fill the gap—within a few generations individuals and cultures whose psychology or rules moved them to have more kids—eg religious fundamentalists, highly impulsive, strong maternal or paternal drives and strongly patriarchal cultures (Looking around world places that treat women like shit seems to have higher fertility) would steadily grow to dominate the gene pool—seeing a greater number of kids being born.

Foyle Aug 2, 2023, 6:35 AM
2 points
−11
on: My current LK99 questions
purported video of a fully levitating sample from a replication effort, sorry I do not have any information beyond this twitter link. But if it is not somehow faked or misrepresented it seems a pretty clear demonstration of flux-pinned Meissner effect with no visible evidence of cooling. [Edit] slightly more detail on video “Laboratory of Chemical Technology”

Foyle Jul 26, 2023, 9:47 AM
12 points
0
in reply to: johnswentworth’s comment on: The First Room-Temperature Ambient-Pressure Superconductor
Seems quite compelling—most previous claims of high temp superconductivity have been based on seeing only dips in resistance curves—not full array of superconducting behaviours recounted here, and sample preparation instructions are very straight forward—if it works we should see replication in a few days to weeks [that alone suggests its not a deliberate scam].
The critical field strength stated is quite low—only about 25% of what is seen in a Neodymium magnet and it’s unclear what critical current density is, but if field reported is as good as it gets then it is unlikely to have much benefit for motor design with B² dependent torque densities <10% of conventional designs, unless the applications are not mass/cost sensitive (wind turbines replacing permanent magnets?).
Meissner effect could be useful for some levitation designs (floating houses, hyperloop, toys?) Likely some novel space applications like magnetic sails, perhaps passive magnetic bearings for infinite life reaction control wheels and maybe some ion propulsion applications. But lightly biggest impacts will be in digital and power electronics with ultra-high q inductors, higher efficiency transformers, and maybe data processing devices.
It might be transformative for long distance renewable power distribution.
[Edit to add link to video of meissner effect being demonstrated]
Meissner effect video looks like the real deal. Imperfect disk sample is pushed around surface of a permanent magnet and tilts over to align with local field vector as gets closer to edge of cylindrical magnet end face. Permanent magnets in repulsive alignment are not stable in such arrangements (Earnshaw’s theorem) - they would just flip over, and diamagnetism in conventional materials—graphite the strongest—is too weak to do what is shown. The tilting shows the hall-marks of flux pinning working to maintain a consistent orientation of the superconductor with ambient magnetic field, which is a unique feature of superconductivity. No evidence of cooling in video.
If this is not being deliberately faked then I’d say this is a real breakthrough.