HNHacker News
TopNewBestAskShowJobs

dgacmu

8,057 karma · joined February 12, 2013

CTO and founder, Enriched Ag. Computer science professor at Carnegie Mellon University. Ex- university of Utah, MIT, Google Brain.

http://www.cs.cmu.edu/~dga/

Of the firm belief that distributed systems is the coolest area ever, followed by computer science in general. Yes, I'm a bit biased.

submissionscomments
dgacmu··on Sodium-ion "salt" batteries will revolutionize electric-vehicle and grid storage
You seem to be turning this into a discussion of vehicles, but the post you're replying to was purely about batteries. I get it - ICE cars suck. But speaking for myself, I waited until I could get LFP batteries before adding a lot of lithium to my house. I deemed the fire risk of NMC not worth the benefits. I now have about 15kWh of LFP and I'm happy as a clam.
dgacmu··on A better way to tie your gym shorts. (Or any drawstring) [video]
Strongly agreed, and thank you - TIL about the truckee hitch, I've always been using trucker's hitches and now I'm going to go try to reprogram myself.

I'd add a butterfly loop to this list for those times you need to add a tie-in point to the middle of a rope for whatever reason.

dgacmu··on CarPlay Is Additive
Consumers having a preference is not ok?

I feel the same way about Android auto. I refuse to be locked into some terrible, never updated or expensive subscription vendor nav unit. I have a phone. I want to be able to use it.

dgacmu··on Understanding lattice risks: Many differences between marketing and reality
Filippo Valsorda doesn't work for Google. He left there several years ago. He runs an independent "buy contracts with me to support development of open source cryptography software" thing.
dgacmu··on Understanding lattice risks: Many differences between marketing and reality
I have no horses in this race and lack the expertise to judge, but this is part of a larger fight and djb seems to be on the other side of a lot of good cryptographers:

https://mailarchive.ietf.org/arch/msg/tls/ZBpcicZX1Dxnam2gMa...

dgacmu··on Nano Banana 2 Lite
Accepting 100% that it should be in some way deemed unacceptable (socially or legally) to fake what an apartment actually looks like, I did find using an image model really helpful in making design choices for my bathroom remodel. Mostly about whether to tile certain things where we couldn't quite visualize ourselves what the effect on the entire space would be.
dgacmu··on Claude Sonnet 5
One of the best queries I've done with an LLM recently was: Create a plan for improving the robustness and resilience of this code, particularly to untrusted inputs.

Gemini wouldn't do a security audit. But it came up with a great set of mitigations and identified an extant XSS flaw in the process of improving robustness.

There's an awful lot of good that can come from proactive, defensive use of LLMs. I realize there's also a lot of pain when the difficulty of exploit finding drops suddenly, but in the long term we may all benefit from the defensive side of this.

dgacmu··on Exercise intensity influences body composition in healthy older adults (2025)
Read my comment more carefully if you think I'm describing the study instead of providing more general context for why people might find this particular study interesting.
dgacmu··on Exercise intensity influences body composition in healthy older adults (2025)
I'm giving context for why a study of higher intensity cardio -in older people- might be interesting to people, and why it's not just "exercise good"
dgacmu··on Exercise intensity influences body composition in healthy older adults (2025)
Mostly because functional strength is useful and keeps you alive. major goals as you age are avoiding falls and being able to continue doing things for yourself. Strength fits that bill pretty well (and it also improves fat free mass).

And on a slightly more technical note, recovering from higher volume becomes harder as you age, so focusing on a smaller number (5ish) of reps at higher weight gives you adaptation without quite as much stress.

But I should be clear, when I said real lifting, I don't mean to exclude any form of well calibrated progressive overload, whether that's strength focused or hypertrophy focused. I do mean to exclude the "go to the gym and lift a 10 lb weight the same number of reps each time" BS

dgacmu··on Exercise intensity influences body composition in healthy older adults (2025)
No, but past recommendations for older adults (note that the average age in the study was 72 years old) were towards "gentle" or moderate exercise. We're seeing a shift now towards recommending real weight lifting and higher intensity as we age. ("Real" -> closer to powerlifting in terms of goals and methods)
dgacmu··on Qwen 3.6 27B is the sweet spot for local development
That's incorrect, I have one on my desk right now. They've stopped selling it now, but I got one a year and a half ago:

> Apple M4 Pro chip with 14‑core CPU, 20‑core GPU, 16-core Neural Engine 64GB unified memory 2TB SSD storage 10 Gigabit Ethernet Three Thunderbolt 5 ports, HDMI port, two USB‑C ports, headphone jack Accessory Kit $2,649.00

dgacmu··on TOP500 at ISC’26: We have a New Number 1 Supercomputer
That's not nearly as true when you look at AI training clusters. They're basically supercomputers but without an FP64 focus.

(These are the systems to which GP was referring at Google.)

dgacmu··on 45°C cooling design cuts data center water use to near zero
Sorry! Power Usage Effectiveness. It's a measure of the overhead of delivering power to the servers (or motherboards, depending on how you measure it), taken to include all other datacenter power overheads: AC/cooling, UPS losses, and sometimes transformer/rectifier losses.

A PUE of 1.2 means that for every watt of power you're using in a server, you're using 1.2 watts drawn from the grid (approximately). Those extra 0.2 watts are being spent on cooling and other things.

dgacmu··on 45°C cooling design cuts data center water use to near zero
As someone else noted - it's effectively zero compared to the water evaporated in evaporative chillers.

But it's still misleading. The major source of water use in datacenters, by _far_, is the water used in power generation. This improves PUE, which reduces power draw, but the savings are almost certainly under about 20% given that many modern datacenters already operate at a PUE of under 1.2. So if you're running on coal or gas, you're still consuming quite a bit of water indirectly.

Now that said again - the water consumption part of this equation is generally overhyped. The power draw is the problem, as are the really bad temporary hacks to the power problem (e.g., what x.ai is doing with "temporary" gas turbines).

dgacmu··on Extreme Heat conference cancelled due to extreme heat warning
You can do a perfectly good, very efficient mini split for USD $5k. Avoids the leaks of window of portable units. And if you're feeling fancy you can get it as a cold climate heat pump. They're great options for retrofitting - can do multiple indoor air handlers, etc., for far less than $50k
dgacmu··on Job application asked for my SAT scores
Speaking as a professor: the filter is really helpful. Having students struggle for two years in a program they're going to fail out of is terrible. I've seen it happen - and I saw more of it happen for the years CMU also stopped requiring the SAT.

The SAT is a very imperfect measure but it turns out a lot of the others are even worse.

dgacmu··on Finding the Best Dog Treat with Statistics
And so useful when you don't feel like scraping the dishes. ;-)
dgacmu··on FDA advisors unanimously vote to approve Moderna's mRNA after agency drama
Both VRBPAC and CDC's ACIP did remarkably good work during the pandemic under enormous pressure - what specific complaints do you have? They got safe and effective vaccines out there fast, did a really reasonable job of balancing the cost benefit w.r.t. side effects of both j&j's and the somewhat elevated myo/pericarditis risk from moderna's vaccine in young men, etc.

Be specific?

(I have complaints - I thought Dr. Meissner was wrongheaded about opposing allowing EUA to extend to pediatric populations. But the committee as a whole functioned well and balanced the urgency of the situation with the need for voluminous data.)

dgacmu··on How many of the 170k English words do you know?
I think you mean it's lognormal, at least if we're discussing native English speakers or comparing those with similar amounts of exposure to the language.

(The median English speaker almost certainly knows several thousand words, or word stems to avoid duplication. But the number who know all words in the tail is exceptionally small.)

dgacmu··on Show HN: Are You in the Weights?
From one person who's completely not-anonymous on HN to another: Why? Do you worry about the association of your IP address with your real name by yet another site other than any place you've made purchases from or signed up for an account from?

(I'm asking seriously, as I can see some risk to having that linkage more public, but given the rate with which services holding PII are compromised and my own personal rate of receiving notices of "oops, we kinda sorta leaked everything about you, here, have more free credit monitoring", I assume almost all of this is available already.)

dgacmu··on Law Enforcement's "Warrior" Problem (2015)
The problem for people like that is that they're working in a system that rewards and mandates bad behavior. You want your traffic tickets to stick so you can make quota? Ticket more black and poor people. They won't contest the tickets as much. Drug arrest and conviction quota? Find people who can't afford an attorney, they'll get an overworked PD and likely take a fast plea bargain.

Good people are responsive to the incentives we've created for them also.

dgacmu··on Nvidia is proposing a beast of a CPU system for Windows PCs
He's not a loser; he's done some really fun work that many people use daily. I've used his range mapping trick in multiple projects/papers. It's elegant.

It sounds like he's gotten bad advise about how to market himself /or/ this is being marketed to people who have bigger checks to write and whom he believes will be responsive to this kind of marketing. As an academic, it rubs me very wrong - I think it's detrimental to the field when we get into h-index stacking contests or citation count comparisons. But I don't know what incentives he's responding to, which seems important for putting this stuff in context.

(as an aside, it turns out that polars + fastexcel is about 10x faster than pandas + openpyxl for searching that dataset, if anyone else is curious what he was actually talking about. :)

dgacmu··on Ask HN: What was your "oh shit" moment with GenAI?
I suggested to a masters' student that a problem we were working on would benefit from analyzing it mathematically. He brought an incorrect solution the next time we met, and on a whim, I asked Gemini to do it. Gemini got it right. I started looking for more ways to use it after that.
dgacmu··on Gemma 4 12B: A unified, encoder-free multimodal model
Ah! Good question: Google's non-open-weights models (Gemini, etc) have almost always outperformed on image recognition tasks compared to any other models. I use a mix of in-house and Gemini for image classification tasks for $startup. No other models have done as well, and I had hoped that some of that would spill over into their open source models. It does to a degree - bigger Gemma models are okay.
dgacmu··on Gemma 4 12B: A unified, encoder-free multimodal model
I was excited about this until I fed it one of my local test problems: coin identification. I then spent 10 minutes arguing with it that a photo of a 1998 washington quarter was not, in fact, a Morgan Silver Dollar. I mean, I wish it was.

It went into a crash loop on a british columbia 1 dollar coin. This happened with both Q4_1 and Q8. Maybe I'm holding it wrong or it's just really bad for this task.

In contrast, gemma4 gets the british columbia coin right though it also mis-identifies the quarter. gemini 3.1-flash-lite nails them both.

Was getting about 50 t/s output on a 3090 with Q8 which seems ok.

dgacmu··on A new way to build chips: Sequentially stacking silicon to extend Moore's Law
But it does, at least to some degree: For the cpu (or brain) you want high density to minimize the latency between components. For the heat sink, you want high surface area, which you can actually do to some degree three-dimensionally, particularly when you have active cooling. Look at a typical heat sink with a fan attached to it -- it has some depth, because that depth allows more heat to be transferred to the air to it by increasing the surface area exposed to the air. Lungs do the same thing, and they do function as part of our cooling system. So if you have a way that the flow of the exchange medium is not limited by the external surface area of your heat exchanger (a fan, a pump, a diaphragm), you can go pretty far.
dgacmu··on The newest Instagram “exploit” is the goofiest I've seen
I'm probably out of date, but Google's advanced protection at one point did account recovery via postcard to your home address. High latency but pretty good as a fallback.
dgacmu··on WH proposes rules giving political appointees final approval on research grants
One of the explicit goals of the NSF is to train the next generation of scientists. Part of that is making sure that you're creating a rich pipeline of people who are going to do innovative things. Broadening participation is much more about things like getting more (usually younger) people from all walks of life interested in joining your field. Which is basically an unmitigated good -- first, the obvious advantage that having more people who want to be in a field is good for it from the perspective of choosing the best folks. And second, the less obvious but perhaps more important thing that people with different perspectives often end up thinking about problems differently. It's not nearly as helpful to have 1000 people all focused on chasing the same problems with the same toolbox of solutions as it is to have 1000 people focused on different problems with different ideas of how to approach them.

I say this as a professor at a top computer science department. I have _never_ felt limited in my ability to collaborate with the best folks in my area. Ever. I do! And it's great! And I also believe strongly it's important to make sure we are growing those next generations of amazing people, because the thing that makes research awesome is working with them.

dgacmu··on FBI arrests CIA official with $40M in gold bars in his home
But then you'd have to keep stealing more gold bars for the upkeep, and who has time for that?
← PreviousPage 5 of 34Next →