Don DeRosa
165 posts
Don DeRosa
@EonixBatt
Battery Scientist | CEO of Eonix | Attends way too many Vols games
Knoxville, TN Katılım Aralık 2021
226 Takip Edilen139 Takipçiler
Excited to share our preprint: Molecular Representations for Large Language Models.
We show that LLMs struggle with existing chemical formats, and that our new MolJSON representation substantially improves performance.
English
OK wait...
perplexity actually runs Karpathy's LLM Council natively
they call it "model council."
multiple models answer your question independently, then a synthesis layer combines the best reasoning from all of them into one response
example: you ask "should i price my product at $97 or $197"
instead of one model giving you one answer, 4+ models each make their case separately.
then perplexity merges the strongest arguments into a single verdict
here's how to enable it:
1. open perplexity on web
2. click the model selector
3. select "model council"
4. ask your question
based perplexity.
this feature now makes it my most trusted LLM for queries out of the box
all ai labs should implement some version of this
Ole Lehmann@itsolelehmann
English
@itsolelehmann @tradercassandra I think with frontier models and the right agent structure its likely already possible. We use 100% local models, so it might be a year or so more until we can remove our human in the loop. Need Zuck to step up and release something better soon.
English
@EonixBatt @tradercassandra do you see this change in the near future?
English
The 40% cost reduction is amazing but still kind of undersells it
The real number is the time compression.
A human researcher might test 20-30 combinations in a good month. This system tested 6,000 per iteration.
(Which is roughly 150 years of traditional lab work compressed into a few weeks, if you want to feel something about that)
Drug discovery, materials science, synthetic biology, basically any field where the bottleneck is "we need to try thousands of things to find what works" just got its timeline crushed
The second-order effects of this will be insane
OpenAI@OpenAI
We worked with @Ginkgo to connect GPT-5 to an autonomous lab, so it could propose experiments, run them at scale, learn from the results, and decide what to try next. That closed loop brought protein production cost down by 40%.
English
@LimitingThe @AustinMeyers_ @SawyerMerritt It has lithium metal specific energy and supercap lifetime . . . Feels like a Nikola moment
English
@AustinMeyers_ @SawyerMerritt That's weird, I didn't get a notification for this
Sounds like BS to me 😁
I would absolutely love to be wrong
Generally cycle life and energy density work in opposition to each other
English
NEWS: The world's first production ready all-solid-state battery vehicle has been unveiled, a fully electric motorcycle.
• Starting price: $30,900 ($44,900 for top trim)
• Range: Up to 370 miles (top trim)
• 80% charge in 10 minutes
• Native NACS port
• Peak charging speed: 200kW
• 33.3 kWh battery pack
• 0-60mph: as little as 2.5s
• Energy density of 400 watt-hours/kilogram (vs ~300 Wh/kg in Tesla's new 2170 cell in the Model Y Performance)
• 100,000 cycles, without needing to limit charging to 80%
• Retains over 99% of its capacity at temperatures as low as -22°F (-30°C) and as high as 212°F (100°C)
• Battery will not ignite if it’s damaged, and doesn't use any rare earth materials
• Modular architecture
• Up to 201HP
• 124mph top speed
• Deliveries start in the US in Q1 2026. Orders open today.
The specs are impressive on paper, but whether all this is actually true will need to be confirmed when deliveries start in a couple months.
More images in thread below, including of the native NACS port:
English
@the_antrik @LimitingThe A 21700 being charged from ~20% to 80% at 2C benefits greatly from multitab. Tabless might be overkill given the manufacturing complexity, but multi-tab yields significant benefits. Single tab cells can have a resistance that is 4X higher.
English
@EonixBatt @LimitingThe No, the opposite would be crazy. The whole point of multi-tab is keeping cooling performance while increasing diameter. 21700 simply has no need for that.
English
A lot of people are asking if the 21VB cell in the North American Performance Model Y from Panasonic Kansas has higher energy density than the battery cells produced at Panasonic Nevada.
The answer is yes.
Based on the testing and reviews I've seen so far, it has between 6% and 10% greater energy density.
If that's the case, the total energy density of the new Panasonic 2170 is between 290 and 300 Wh/kg.
This is in the right ballpark for the newest generation of cells that are coming out in the next couple of years.
As usual, Panasonic is at the head of the pack in terms of energy density when it comes to mass manufactured high nickel battery cells.
I'm betting with the next version of the cell (21VC) they focus heavily on charging speed while trying to maintain the same energy density as 21VB.
The Limiting Factor@LimitingThe
Further information has come through: 1) They're pushing hard on the 21VB to exploit the remaining IRA tax credits and other benefits of domestic manufacturing (e.g. tariff avoidance). It's a win-win for Tesla and Panasonic. 2) Earlier this year I said that Tesla and Panasonic were working on a higher energy density/faster charging version of the 2170. That cell is different and referred to as the '21VC.' It will be manufactured starting in Giga Nevada and it's expected to start production in around a year (so maybe won't show in vehicles until 2027?) I'll share my post on that cell just below this one.
English
@LimitingThe They are at least multi-tab then right? Single tab would be crazy
English
@EonixBatt No word yet
But, not needed for the 2170 because it's so small.
So, unecessary complication in my view. Easier to just use existing equipment and de-risk the production ramp.
English
@EVCurveFuturist We haven’t tested a CATL sodium ion cell yet, but every sodium ion battery we have reviewed so far actually seems to be more explosive than conventional Li-ion. I am curious to hear the pitch on how they are safer as it hits the market.
English
Sodium-ion isn’t a #lithium replacement — it’s market segmentation.
It targets use-cases where cost, safety & cold-weather performance matter more than energy density: entry-level EVs, fleets, battery swapping, some stationary storage. Those were never the core drivers of lithium demand.
Lithium still owns the heavy hitters: long-range BEVs, fast-charging platforms, large packs (75–100+ kWh), grid-scale BESS, and now AI/data-centre power. Physics hasn’t changed.
Even CATL frames this as a dual-track future: sodium alongside lithium, not instead of it.
Here’s the counter-intuitive bit markets miss: sodium can increase lithium demand. Cheaper entry EVs widen adoption, accelerate the S-curve, and expand total battery kWh deployed. Lithium demand is driven by energy, not unit count.
Sodium trims the edges. Lithium still owns the core.
carnewschina.com/2025/12/28/cat…
English
@MaxPatten @electricfelix I think they have come a long way in mitigating alot of the risk of accelerated degradation during fast charging but 40 intermittent cycles over the lifetime of the pack at even V4 Tesla charging rates will wear it pretty bad which for us is five road trips (4 stops).
English
@EonixBatt @electricfelix I trust that you know batteries better than me but I also don’t think ultra fast charging is a stunt. 400 kW is already proven in mass market vehicles and quite compelling.
English
The amount of cope in the comments is shameful. If western automakers took charging seriously years ago this wouldn’t be such a shock. High voltage, new battery chemistries, and scale is a crazy thing.
Balaji@balajis
This is the EV flippening. Electric vehicles now fill up roughly as fast as gas. The new BYD chargers add 400km of range in just five minutes.
English
@MaxPatten @electricfelix Yeah i realize that, i also realize the limitations of the battery chemistries they are using and touting. That will mess their battery after a few charges, but hey I am not saying dont buy it. Go for it and see if their warranty is better than the Chevy Bolt.
English
@EonixBatt @electricfelix Uh huh. So why are they warranty-ing it and selling it in a car then?
English
@MaxPatten @electricfelix What is the warranty? How are they planning on conducting a SOH check under warranty? What warranty caveats are built in to void it? Is there an independent company that evaluates the capacity of the pack or do they have the opportunity to diesel-gate it themselves?
English
@sebastienroblin Given the strain on supply chains and production, seems like mobility kills are good enough!
English
At say $15,000, still cost-advantageous compared to $1+ million tanks and $10,000s anti-tank missiles. But surging 30-40 FPV drone attacks--each individually piloted--to kill one vehicle at a specific place and time is a major effort! Easier if tank is immobilized of course. 2/2
English
Very interesting thread broadly, just to highlight a couple tech/ops bits: 1) systematic deployment of unmanned ground vehicle platoon/company/battalions in Ukrainian formations! 2) average of 30-40 kamikaze FPV drones to destroy an assault tank. 1/2
Michael Kofman@KofmanMichael
An update on the war following a recent trip. Ukrainian forces are holding, but the situation has worsened since July due to mounting offensive pressure. Here I cover some of the negative and positive trends, along with the salient dynamics at the front. Long thread. 1/
English
Don DeRosa retweetledi
We blew lithium ion batteries up in the name of science.
@EonixBatt has automated the R&D and testing of batteries to get new, safer chemistries faster.
All on Part Two of our tech road trip through Tennessee, which I can never spell right on the first try. Thx to the folks @UTKnoxville for being so generous with their lab
Full episode in the comments.
And massive thanks to @brexHQ for letting us bring this frontier science to the world
English
Check out @corememory ‘s tennessee road trip! As a long time fan of Hello World, it was a privilege to give @ashleevance a tour of Eonix’s automated materials discovery lab in Knoxville next to @Vol_Football’s Neyland Stadium. Kid Rock and @WhisperAero were in the earlier episodes too so check them out as well, they were great!
Ashlee Vance@ashleevance
Went to Knoxville and saw some very weird things. A machine that can build infinitely long parts - think entire bridges on site. And a dude changing how we make lithium ion batteries @EonixBatt Full episode here and will drop YouTube link in the comments. This is Part Two of our Tennessee tech road trip. Thanks to @brexHQ for making our exploration of the future possible Chapters 00:00 - Intro 00:45 - A Weird New Machine 05:14 - Making Better Lithium Ion Batteries 09:10 - Let's Make One Go Boom
English