HNHacker News
TopNewBestAskShowJobs

dgacmu

8,038 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 AI-generated posters don’t have to be horrible
One of my best memories from an art class in high school was that a student organization came into our class and asked if any of us would be willing to design a poster. I ended up designing and silk-screening their poster for them and it was a fantastic experience.

I can totally see how doing more than a couple would start to feel like work, but having one or two experiences like that was awesome.

dgacmu··on Small programming tricks
Also in academia but I'm old and my real work is done withy students. I'm coding by hand much less than I used to (boo) but solving more problems with code myself than I used to (yay). Having an llm do things turns out to be similar to having a student do them but with fast turnaround and it's fine to have them work on your home automation system. ;)
dgacmu··on Bad benchmarks and evals: Senior SWE-Bench, napkin math, and winter tires
I agree, I have the same problem, that's why I'm sticking with the crossclimates year round also -- we only drive about 5,000 miles per year, so my tires die from age well before they would wear out from mileage.

I assume we're both making approximately the same optimization but you're going for better snow & ice performance at the expense of some summer and wet stopping distance, and neither tradeoff seems obviously wrong depending on our driving situations.

dgacmu··on Bad benchmarks and evals: Senior SWE-Bench, napkin math, and winter tires
I dunno about this - I leave my crossclimate 2 aw's on year round and they're pretty fantastic in snow. We had a pretty good winter this year (44" from dec-feb) and they just worked. I didn't go meandering around any mountains on them, mind you, but they're so far ahead of most typical all-season tires it's almost not fair to compare. For -most- lazy people who don't want to swap tires or rims, they seem a better option than leaving true winter tires on year round. Obviously, there are exceptions depending on where you live.
dgacmu··on Base84 deserves a place in file names
Yeah, I'm sympathetic to their desire for compression and efficiency but base64URL is way more shell-friendly.

I quite like base58 for human-presented identifiers also. I tend to use it instead of hex for having ~8-16 character IDs of things.

dgacmu··on I Added a Non-Wi-Fi Mitsubishi AC to Home Assistant
But with most houses, the AC will cycle on and off throughout the day anyway - raising the temp while you're gone will often result in less cycling. Especially if you also increase the hysteresis range while out. So you get a triple efficiency win, from better AC efficiency, less AC motor wear, and less heat loss to the outside.
dgacmu··on LRU is harder to beat than the KV-cache papers suggest
"That's not a result, that's a broken harness, and it's worth publishing because I expect it to be common."

That's Claude. But I agree and am also glad the author, uh, clauded this up. :)

dgacmu··on Ask HN: Can we please limit the AI news flood?
On the brax thing - the "article" is a youtube video. There's a pretty reasonable amount of pushback here against video-only content where folks can't rip through quickly to figure out how relevant it is to them or if it's just a vapid ad.
dgacmu··on DeepSeek v4.1 Flash
Bandwidth is really cheap in bulk. You can get a 100 gigabit internet connection for about $10,000 a month. If you were to somehow keep that saturated 24/7, you'd move about 30 petabytes in a month, so your per gigabit cost is only $0.0003.

Realistically, if you had it 5% utilized, those million tokens would cost you about $0.0000062, which is pretty insignificant compared to what they charge you. (Assuming one byte per token, ignoring compression)

dgacmu··on I Connected My Withings Body+ to Home Assistant with an ESP32
esp32 bluetooth proxy: There's a standard esphome component that just turns any esp32 with appropriate h/w into a BTLE proxy: https://esphome.io/components/bluetooth_proxy/

I'm just using a cheap ESP-WROOM-32 but the recommended way to do it is with an wired ethernet-based ESP32 so you don't have interference with the wifi and btle. For a scale this works well enough. I haven't yet received the athom smart plug so I can't recommend it yet. Should be here in a week or so.

Once you've got the proxy you can just use the eufylife integration: https://www.home-assistant.io/integrations/eufylife_ble/

To setup the esphome bluetooth proxy you install esphome, and then create a yaml file that you'll use to configure the board, like this:

   esphome:
     name: bluetooth-proxy
   
   esp32:
     board: esp32dev
     framework:
       type: esp-idf
   
   # Enable logging
   logger:
   
   # Enable Home Assistant API
   api:
     encryption:
       key: "YOUR_KEY_HERE"
   
   # Enable OTA updates
   ota:
     - platform: esphome
   
   # WiFi configuration (matching the settings from your other devices)
   wifi:
     ssid: "YOUR_WIFI_SSID"
     password: "YOUR_WIFI_PASSWORD"
   
     # Enable fallback hotspot (captive portal) in case wifi connection fails
     ap:
       ssid: "Bluetooth Proxy Fallback"
       password: "proxyfallback"
   
   captive_portal:
   
   # Enable ESP32 BLE Tracker
   esp32_ble_tracker:
     scan_parameters:
       active: true
   
   # Enable Bluetooth Proxy
   bluetooth_proxy:
     active: true
Once that's done, plug in your board to your system and use esphome run bluetooth-proxy.yaml (you might have to find the device it's showing up as in tty.usbserial or whatever).
dgacmu··on I Connected My Withings Body+ to Home Assistant with an ESP32
If you have a eufy scale, you can do this with no LLM needed - just set up an esp32 Bluetooth proxy, add it to home assistant, and then use the EufyLife integration someone already made. I did this recently with my scale to avoid having to sync it to my phone, go to the cloud, fetch data from the cloud, etc., and it's working fantastically. I have a little Home Assistant integration that pokes scale updates to a Python script running locally that provides a UI I like.

Huge fan of running it this way.

(Yesterday I ordered a more commercial, polished-looking esp32 Bluetooth proxy integrated into a smart plug - athome brand - but it's the same functionality, just in a nicer package than having a USB thing dangling from my outlet)

dgacmu··on Research acceleration: The view inside OpenAI
Indeed, many programmers might pattern match to repetitive stress injury and think of their brushes with carpal tunnel syndrome. :)
dgacmu··on Artificial Analysis Intelligence Index v4.2
It does and that one also feels kind of saturated for measuring the most advanced models - opus, Gemini, astra, sol, fable, glm, kimi all scoring within statistical noise of each other. (74 +-3% down to 69% +-5% for kimi).

It's still providing strong discrimination between weaker models.

dgacmu··on Artificial Analysis Intelligence Index v4.2
I quite wish they'd move to terminal bench 4.0. 2.1 is saturated - there is no world in which Gemini 3.8 Flash is producing better code than Astra or Fable, as the 2.1 results might suggest. The 4.0 results differentiate these models much more effectively.

(2.1 is useful for knowing they can all one-shot straightforward scripts, of course.)

dgacmu··on IBM Bob
I think they started on a long slide starting in the late 90s when they tried to couple research more and more tightly to product deliverables, and then as they pivoted to being a services company, including selling off non-mainframe hardware to Lenovo, to coupling r&d to, uh, I guess services revenue, whatever that means. They had some research layoffs in 2008, at the same time that MSR and Google were really blossoming in research spend and perception, and I think they kind of got left behind.

I had a former student go there around 2008 and it seemed like the vibes were starting to sour around then. They maintained some awesome researchers in, e.g., databases, but I think the "make sure you are showing closely how your research directly contributes to services revenue" thing was strong by then and somewhat ruinous for how pleasant a work experience it was for many.

dgacmu··on IBM Bob
I laughed when I saw this; my 14yo looked over to see why I was laughing and asked "what's IBM?"

indeed, child, indeed.

dgacmu··on Scientists Find Worms in Almost Every Fish in a Tuna Haul – 96% Were Infected
It doesn't, really - this is a known hazard. You want fish that was flash-frozen very rapidly after harvesting. It's tastier and it kills (these types of) parasites. Which is basically what you're going to get at any decent sushi place.

Never too old to try something that is incredibly delicious. :-) My mom (born 1943) introduced me to sushi. Can't blame generation there.

dgacmu··on Evidence of Fraud in an Influential Study About Procrastination
It should - one of the papers that Data Colada reported on about Gino was co-authored with Dan Ariely: https://datacolada.org/98
dgacmu··on Benchmarking Pocket-Scale Inference
No - DRAM is made on a different process than compute. Apple uses dram made by the usual dram manufacturers.

The dram shortage is a combination of high demand + shifting manufacturing to HBM, which consumes more wafer capacity than DDR5

dgacmu··on Iceland votes on whether to restart talks on joining EU
Those not familiar with it might find the history of the Cod Wars interesting (and very relevant to this discussion): https://en.wikipedia.org/wiki/Cod_Wars

(literally, Iceland having several decades of slightly-military-involved conflict with the UK about fishing near Iceland. Iceland won.)

dgacmu··on "No way to prevent this" say users of only language where this regularly happens
It's a pun on The Onion's repeating headline about gun deaths in the US ("only country where...") [1] . I thought it was pretty clear before I clicked it it was going to be talking about security vulnerabilities in C, but I live close to that world so I might be a biased audience.

[1] https://en.wikipedia.org/wiki/%27No_Way_to_Prevent_This,%27_...

dgacmu··on OpenAI Jalapeño: Better than Nvidia Blackwell
That's only half the problem. OpenAI is contractually obligated, if you will, to believe that models will continue improving at an impressive rate for the foreseeable future (otherwise their valuation makes no sense).

If you believe that, then you should expect to get Sol-level performance out of a Luna-cost model within six months or a year. If you have a system with the weights baked in, that means you're going to end up serving that Sol-class model several times more expensively than it will take someone who comes along a few months later. (such as what recently happened with DeepSeek's update.)

And under that assumption of continuing advancement, baking things in doesn't make sense in general - it's a play you'd make if you think things are slowing down a lot. Which may be right but it's not OpenAI or anthropic's play.

dgacmu··on Scrap (2006)
I don't think he's mocking at all - the Pittsburgh scrapping culture is fascinating. (I've lived here for 20+ years now). I throw stuff on the curb all the time and like magic, it disappears before trash day. I've also picked up a bunch of stuff from neighbors - office chairs, desks, a really nice stash of PVC pipe I used to build a grey water system for my washing machine, etc. it's like a weird little no-internet freecycle. Everyone just knows it often works and goes along with it. There are people who focus on higher value items, people who grab metals, you name it
dgacmu··on Scrap (2006)
Most likely you're seeing the causality in the wrong direction - poverty creates a strong incentive to go for immediate benefits rather than absorb the risk of one that is delayed.
dgacmu··on $12B of US ratepayers' money wasted on a modeling mistake in PJM
But that's why they're arguing that repricing the entire generation fleet is bad compared to having a different auction for new capacity. You don't need the same incentives to keep an existing, profitable plant online as you do to invest in a new plant.
dgacmu··on The Case Against Formal Verification, 50 Years Later
Facebook runs a number of quite complex internal distributed systems - databases, caches, proxies, etc. all of these are amenable to various forms of formal verification, and verifying them is the kind of thing that helps prevent outages and data loss.
dgacmu··on Claude: System Prompts
Your second two had high comments to upvotes, which tends to get articles downranked more quickly. It may simply be that the stuff you're posting is generating disagreement without corresponding upvotes.
dgacmu··on Xorshift Generators
Will do. I'm on vacation right now and losing my laptop for a few days, but seems worth trying. My recollection from the RNGs a decade ago (I'm dating myself) was that the AES approaches had higher latency but were quite decent, though slower than PCG. Curious how that's evolved.
dgacmu··on What happens when an LLM never sees material beyond fifth grade?
Oh, that's interesting - good point, since it's filtered and not trained from scratch. My prior would be to assume it's just bs'ing as LLMs usually do but it seems worth exploring.
dgacmu··on What happens when an LLM never sees material beyond fifth grade?
I prefer my 8yo's answer about quantum entanglement, asked just now: "I don't know. How would I know? It's not a thing!"

Even an 8yo has better metacognition, it seems. :-)

← PreviousPage 3 of 34Next →