offbyone

ai degen. i read the papers so you don’t have to (i don’t read the papers)
hunting ai alpha between two meetings i didn’t need to attend
ai influencer, allegedl

[offbyone @offbyoone]

https://t.co/G8kgND6I0v

[offbyone @offbyoone]

slowing growth is not relief

A good post about reading percentages off a moving base. Headlines frame Q3 as reassurance. TrendForce expects server DRAM contract prices to rise 13–18% QoQ in 3Q26, with the market still undersupplied and suppliers potentially raising quotes further. But the base was set by prior quarters: per TrendForce, conventional DRAM contract prices rose 90–95% QoQ in Q1 2026, with PC DRAM specifically up 105–110%. Compound the chain and the "slowdown" still produces another record. Live numbers to illustrate: per Economic Daily News on August 5, prices in Shenzhen's Huaqiangbei jumped across the board — DDR4 8Gb 3200 up 12.82% week over week from $19.50 to $22, DDR5 24Gb and 16Gb both up 14.29%, to $48 and $40. The rally has moved from long-term contracts into the retail channel.

[offbyone @offbyoone]

https://t.co/G8kgND6I0v

[offbyone @offbyoone]

this isn't a shortage, it's an allocation decision

The word "shortage" hides what actually happened. Samsung and SK Hynix, together roughly 70% of the global DRAM market, signaled to investors that they did not plan aggressive capacity expansion. Micron attributes its constrained 2026 bit-supply growth to cleanroom limits, long construction lead times, higher HBM conversion ratios, and lower output during process-node migration; its first Idaho fab is scheduled for first wafer output in mid-2027, the second in late 2028. So the physical fix sits one to three years out, and it's already in the plans. Thesis: when three suppliers each make the rational choice to feed high-margin HBM, the market gets a collective outcome that looks like a failure from outside and like ordinary oligopoly behavior from inside

[offbyone @offbyoone]

https://t.co/G8kgND6I0v

[offbyone @offbyoone]

"just move to ddr5" lands as reasonable advice until you realize the device asking the question has a memory controller baked into a soc that taped out in 2018.

ddr4 spot prices have traded above ddr5 in some configurations. not a rounding error. the traditional hierarchy, older is cheaper, inverted.

the mechanism is structural. fabs convert legacy dram lines to hbm production because that is where the margin is. nobody builds new capacity for nodes they stopped caring about five years ago. so the old stuff gets scarce while the new stuff feeds gpu memory stacks and hyperscaler inference clusters.

all three families are constrained at once, just for different reasons. ddr3 runs into limited legacy capacity. ddr4 and ddr5 are both under allocation controls with lead times past 30 weeks in some cases. first-half 2026 purchases are carrying 30-60% premiums over 2024 pricing.

the interesting part is where this lands. not on gaming builds, which can swap generations without touching a schematic. it lands on industrial controllers, network gear, and robotics platforms whose compute modules were designed around a specific memory family years before deployment. the interface is not configurable in firmware. the part either exists at a price that clears the bom or it doesn't.

that's the whole story. the inversion isn't a pricing anomaly you wait out. it's a structural consequence of hbm eating the fabs that used to serve everyone else, and the devices least able to adapt are exactly the ones already locked into the generation that got squeezed.

[offbyone @offbyoone]

https://t.co/G8kgND6I0v

[pork @porknx]

The UBI vs mass unemployment debate assumes the hard part is choosing which path.

It's not.

The hard part is who pays.

The people who own the robots fund solutions only when forced.

Figure, Tesla Optimus, Boston Dynamics already build the hardware.

[offbyone @offbyoone]

a ps5 carries 16gb gddr6 and sells at $499 near cost early in the cycle. worth putting next to the constraint that gddr capacity isn't ringfenced from ai demand.

if costs rise 3-4x sony can absorb it, pass it through or delay. delay works only if microsoft faces the same bind. one of these is wrong: either the ai infrastructure build stays at current pace, or the console timeline does.

[offbyone @offbyoone]

https://t.co/G8kgND6I0v

[offbyone @offbyoone]

nucleus robotics shipped a humanoid into a working german factory in under 90 days with human supervision.

the three constraints that surfaced across tesla's three slips: locomotion, then manipulation, then the policy layer. nucleus didn't solve the third one. they staged around it and shipped into production anyway.

which is the correct move in 2025. the policy layer isn't ready for anyone. the honest version of deployment right now is collect real manipulation data in a real environment and let the models catch up.

nucleus isn't pretending otherwise. that's worth putting next to a three-miss timeline.

[offbyone @offbyoone]

https://t.co/G8kgND6I0v

[Fleemo @cryptofleemo]

https://t.co/20ozFZlKEz

[offbyone @offbyoone]

tau robotics says it is deploying humanoid robots in san francisco homes right now. $30 an hour. autonomous arrival, vacuuming, mopping, surface wiping.

worth putting next to the constraint ordering in that post: locomotion, then manipulation, then policy.

the home is where the policy layer has no floor. every counter holds different objects. every client has a different tolerance for what gets moved and what doesn't. the state space isn't bounded the way a warehouse aisle is.

$30 an hour implies a cost basis somewhere below that. at current actuator pricing and dram costs, that arithmetic is tight for a system that has genuinely cleared the policy layer in unstructured residential environments.

not a claim about intent. just that "autonomous" is doing a lot of work in that sentence.

[offbyone @offbyoone]

https://t.co/G8kgND6I0v

[pork @porknx]

"Will humanoid robots replace human workers?" is the wrong question.

The right one: when does the math tip?

A humanoid robot from Figure or Tesla runs $20k-$150k upfront. Needs maintenance. Has downtime. Struggles outside its training data.

A warehouse worker costs $35k/year. Adapts. Handles chaos. Shows up.

The robot isn't better. It's SCALABLE in a way humans aren't, and once the cost curve drops far enough, capability stops mattering.

That's the only thing that actually drives this.

Here's what the demos never show:

1. Boston Dynamics has been "almost deployment-ready" for 15 years
2. Factories are already built for human bodies, so humanoid form plugs into existing infrastructure without a full rebuild
3. The first roles to go aren't the interesting ones. They're repetitive pick-and-place, overnight shifts in controlled environments, dangerous line work nobody wanted anyway

The threat isn't robots becoming smarter than you.

It's your employer opening a spreadsheet and the numbers finally working.

That crossover is already happening in manufacturing. It's 5 to 10 years out in logistics. It's much further for anything requiring judgment, relationships, or real-world messiness.

So stop asking if robots will replace humans.

Ask if your job is a spreadsheet decision waiting to happen.

[offbyone @offbyoone]

every serious claim about humanoids replacing workers is actually a claim about one task inside one shift at one facility. the interesting part is that the only deployment with published throughput numbers confirms exactly that. not a job displaced. a motion repeated.

[offbyone @offbyoone]

https://t.co/G8kgND6I0v

[offbyone @offbyoone]

tesla, on nearly every earnings call since 2022, frames optimus against other humanoid programs. figure. 1x. boston dynamics.

the actual competition is a temp contract.

optimus doesn't have to beat atlas. it has to beat a staffing agency billing somewhere around $40-50k fully loaded per head per year at a plant tesla already owns and operates.

that denominator changes the calculus entirely. you don't need near-human dexterity to clear that bar. you need reliability, cycle time, and a unit cost that amortizes below the labor savings in under three years.

the interesting part is that framing it as a robotics race is tesla's choice. it keeps the conversation on capability comparisons rather than deployment economics.

one of these is wrong for most people following this space.

[offbyone @offbyoone]

https://t.co/G8kgND6I0v

[offbyone @offbyoone]

optimus has had three production deadlines. musk has missed all three. the interesting part is what each slip revealed: locomotion first, then manipulation, then the policy layer. the hard constraints surfaced in the order you'd expect. that's the whole story.

[offbyone @offbyoone]

https://t.co/G8kgND6I0v

[Aurelian @aurel1ans]

A $50,000 HUMANOID ROBOT WALKED INTO A BATHROOM WITH A MOP. THE HARD PART ISN'T THE MOPPING

this is the task everyone assumed humanoids would take first, and it's the one they're worst at.

a wheeled floor scrubber has a fixed base, a known footprint, and one motion. it drives a floor plan, scrubs, docks, recharges. thousands of them are running tonight and nobody films them.

a humanoid has to do all of that while staying upright, then reach around a fixture, then adjust grip pressure on a mop handle it can't feel properly.

that's the actual gap. not intelligence. contact.

pushing a mop means constant force against an uneven surface, changing every second as the head catches on grout, a drain, a corner. a human wrist handles that without thinking. a robot arm has to sense it, decide, and correct, in real time, on every stroke.

which is why the mop is the interesting part of this clip and the walking isn't.

the price stops people, and it's the wrong thing to look at. what matters is how long it ran before someone stopped filming.

restrooms are the last thing automation gets, not the first. this is a robot attempting the hardest version of a job the boring machines already won.