HNHacker News
TopNewBestAskShowJobs

jhgg

3,448 karma · joined November 4, 2013

currently: Lead of Supercomputing at SpaceXAI, building the biggest AI supercomputers in the world. https://grok.com, https://x.ai

formerly: Principal Engineer of "Core Technologies" at Discord. https://discord.com/

[ my public key: https://keybase.io/jhgg; my proof: https://keybase.io/jhgg/sigs/NfC1mAm4fGJgKsLzYcshaSlYAfvMnxgLUZgee3nWqzI ]

submissionscomments
jhgg··on OpenAI Shoves a Data Journalist and Violates Federal Law
That is unfortunately not how the law works. I am not a lawyer, but I have studied this a bit in passing. It's important to understand the interplay between the scope of the law, private property rights and also reasonableness. Although 26 U.S.C. § 6104(d)(1)(B) mandates that the documents be available for inspection and copying upon request, it does not say anything about allowing civilians to remain on the property beyond the reasonable scope of inspecting these documents. Generally, laws are interpreted in the context of what's reasonable. It's unreasonable to interpret 26 U.S.C. § 6104(d)(1)(B) as allowing someone to stay on private property indefinitely. If they don't have the documents available, and ask you to leave, private property rights apply here and you must leave. The penalty you can seek for non-compliance clearly defined as a $20/day fine, not that you can continue to stay on the property and repeatedly ask for something they do not have a day before a holiday. The Supremacy Clause does not apply here, because there's no inherent conflict. The federal law does not explicitly (or implicitly) state that you can stay on the property after asked to leave, and thus it cannot trump state law. The fact that the initial presence on the property was for a legal, civil purpose does not provide ongoing protection from criminal trespass charges if an individual overstays their welcome or refuses to leave when asked.
jhgg··on OpenAI Shoves a Data Journalist and Violates Federal Law
It's also worth mentioning that failure to immediately disclose said documents are... a $20/day fine [1]. A company being out of compliance does not give a civilian the right to trespass. The correct solution here is to raise a complaint with the appropriate authorities (the IRS), not to continue to remain on private property and have to be forcibly removed by the building's security.

[1]: https://www.irs.gov/charities-non-profits/public-disclosure-...

jhgg··on Privacy is priceless, but Signal is expensive
Paying 2.8m for 20pb of bandwidth is two orders of magnitude more expensive than it needs to be. You pay significantly more using cloud hosting providers rather than dedicated server resellers. I would recommend signal consider just renting dedicated servers from providers like OVH for their voice relays.
jhgg··on Slack is retiring its status account on X
I can't help but wonder why a company retiring a script that syndicates status update to another website is at all newsworthy (or a good enough idea to be worth the possible confusion this might cause to consumers who find it valuable to receive status updates via that medium.)
jhgg··on GPTBot – OpenAI’s Web Crawler
Some would argue that it allows for cutting edge research that could potentially massively benefit humanity. Whether or not that pans out is to be determined, but that is the stated goal. The point then, as advocates for this technology believe is for the betterment of civilization, that by training a neural network on this knowledge will make that knowledge more accessible to the rest of us.

One could of course debate whether or not OpenAI would be the best stewards of that knowledge or aligned with the best interests of humanity. However, it is important to recognize that building a successful business is key to funding the research, H100s aren't cheap. It's also important to note that as with all things tech, price of hardware will go down, and OSS models continue to get more capable every week.

jhgg··on Discord Rolled Out Yubikeys for All Employees
Not replying to something isn't pushing back. Chances are this one feedback post item got lost in the sea of them. If it's any consolation I'll reply to this feedback post saying it's planned and in progress.

As for the QR code login, I built that feature. Although it does offer a venue for social engineering, we've done a lot since launch to ensure people understand that they're logging into a new device using it. From day 1, it's always included red text that said roughly "you're logging into a new device using this" and to "not scan codes that random people have sent you." Of course, some people don't really read. That being said, millions of users every week use the QR code login legitimately, and it's a feature that most other chat platforms offer. It's also very very beneficial when you're using a shared device (i.e. you live in South Korea and visit PC bangs often.)

As well, you'd be surprised, webauthn adoption within companies is not nearly as ubiquitous as you'd think. Shipping out yubikeys around the world during a pandemic was a gargantuan task. Either way, any post advocating for more broad adoption of webauthn and also showing success does the industry as a whole good.

jhgg··on Discord Rolled Out Yubikeys for All Employees
Yubikey support is coming soon to users and is being internally tested on staff right now.

Also, I'm not sure what's given you the impression that there has been any push-back on the feature?

jhgg··on Yelp rebuilds corrupted Cassandra cluster using its data streaming architecture
What is a bit unclear which isn't addressed in the original blog post is how they actually were able to successfully scan over all the data and copy it to the new cluster, in spite of corrupt sstables. CDC would not handle back filling historical data. Would expect that to fail for certain token ranges. Perhaps they ended up deciding to discard them?
jhgg··on Discord is not documentation
Yes because the flag is on your account.
jhgg··on Discord is not documentation
They're called servers because we wanted to pick words that were familiar to what we were originally competing against, which was team-speak/ventrilo/mumble servers.

Picking words people are familiar with in order to convey a concept is a generally smart thing to do. A discord server was meant to be analogous to a ventrilo server. Calling them servers made that idea clear to our users, which helped users understand our product and what purpose a server served. I would say that the term has served its purpose well.

Internally we call them guilds, because having a "server server" is a bit ridiculous.

jhgg··on Type-safe, K-sortable, globally unique identifier inspired by Stripe IDs
I don't think they took offense to the "rita" part.
jhgg··on Real-time messaging
Yeah it's a pretty cool design. Discord's real time system operates on a similar (although much more complicated system) on top of hash rings as well.
jhgg··on Writing a chat application in Django 4.2 using async StreamingHttpResponse
Discord uses websockets on all clients/platforms.
jhgg··on Deep-brain stimulation during sleep strengthens memory: study
If it feels problematic, it's probably because you are missing the fact that there is a control and treatment group of this experiment to control for what you are saying.
jhgg··on The absurd cost of finalizers in Go
These are all reasonable things to do at compile time, however for historical reasons since go did not have generics till recently, this is probably the best they can do without compiler specific specialization (eg what was done prior for maps and slices).
jhgg··on GitHub code search is generally available
It would be really awesome if code search could one day consume LSIF for precise results in its index similar to source graph. The symbol search is good now, but approximate. Having more precise code search by allowing devs to upload LSIF data in their CI pipelines would allow for precise symbol search (go to definition / find usages actually being accurate) and remove irrelevant result.
jhgg··on Show HN: Supavisor – a Postgres connection pooler written in Elixir
cloud sql proxy makes connecting to cloud managed postgres instances pretty simple/easy. honestly, we dont have a "great" reason other than, we set it up a long time ago, and it's continued to work haha.

so pgbouncer connects to cloud sql proxy, which does the connection pooling concern.

jhgg··on Show HN: Supavisor – a Postgres connection pooler written in Elixir
nah we run a pool of common pgbouncers that route to the various postgres instances we run. they also run google's cloud sql proxy as well. it's actually haproxy -> pgbouncer -> cloud sql proxy -> postgres.

everything but postgres runs on one machine. we use haproxy to load balance replica connections across different pgbouncer backends for replicas of the same db.

jhgg··on Show HN: Supavisor – a Postgres connection pooler written in Elixir
at work we run pgbouncer on 4 vCPU (GCP c2-standard-4) machines, at 60% load, and they do ~15k TPS at peak, however we only peak at ~45k connections per node.

however, i think that you may have a lot left on the table in terms of optimizations that can be made. these numbers seem far under the performance of what i know beam to be capable.

jhgg··on I’m a doctor: Here’s what I found when I asked ChatGPT to diagnose my patients
I am curious if GPT-4 would have performed better.
jhgg··on Vector database built for scalable similarity search
I was mulling over the idea of building keyword image search (say, with CLIP based embeddings). However I'm not really sure the cost, or whether or not this is the best solution. Do you have any case studies about large deployments of this software, and what the upper limits of scale might be?
jhgg··on Vector database built for scalable similarity search
Does Zilliz do BYOC?
jhgg··on Arbitrary code execution during compilation – rust
These are optional features. You can decide whether or not you want to use them. No one is forcing you to do these checks at compile time.

My point is, the capability is useful to some people, and there are many other ways that doing arbitrary things at build/compile time can be useful or make things easier. The sqlx example is one of many.

Another usage, is calling out to another tool, e.g. a protobuf code generation tool. That requires the build toolchain interacting with another tool, that would "break the sandbox."

jhgg··on Arbitrary code execution during compilation – rust
Then avoid crates that do such things. Other people however are able to make use of compile time code execution to do some pretty awesome things. For example, a database library sqlx can check all the SQL in your code as being syntactically correct, and also typed correctly against a test database at compile time. A feature that is useful and convenient for users of the library.
jhgg··on How Discord Stores Trillions of Messages
The spark migrator was needlessly complicated in pursuit of being general. It is also much slower.

Ours is quite simple! We take the token range (int64 min -> int64 max), and sub-divide that into N equally sized ranges within the range that span the entire range. We use sqllite to store a database of "work to do" this contains the token ranges, some stats about work done so far, status, and last pagination token (this lets us pause/resume the migration.) We run N many tokio tasks that then just pluck work from the sqllite database, and update the database as work progresses, occasionally snapshotting the pagination token. Each task will insert to the destination database with a configurable amount of concurrency, however will only work on one token range at a time. Each task gets its own "connection" that is not shared with the other tasks to minimize contention. Distributing this workload across machines is very simple. We have a command called "sunder" which just subdivides the sqllite database into N many "shards" that contain 1/N the rows of the original database. We can then just copy these shards to other machines and start the migrator on them. We just spun up 5 chonky VMs on GCP and sundered the database, copied the binary and ran it in screen.

Since we already have the types of the rows described in Rust, we simply just plug the relevant types into the migrator. Everything just works, and you're using the exact same code that's reading/writing the data in the live service to also migrate the data (which further reduces the possibility of bugs). Since the migrator is built on-top of our data service library, we also get all the metrics/telemetry built in as well, so we can monitor the performance of the migration in great detail.

jhgg··on Discord, or the Death of Lore
Do you know of any service that is running at scale out there that stores their data as 5mb plain text files?
jhgg··on Discord, or the Death of Lore
Because it would be obscenely expensive (in terms of computational resources) to search history via grep. Hence why we use an inverted index.
jhgg··on Linode rebranded as Akamai’s cloud computing services
Discord originally hosted on Linode in 2015. Unfortunately we had to move off, as we knew Linode would be the limiting factor in being able to continue to grow. So we chose to go with GCP.

Unfortunately, I suspect this is something that happens quite often for Linode, which is that, their most successful clients will eventually graduate off of them.

This may reposition their business in such a way that their most successful clients stay with them.

jhgg··on Reddit Confirms It Was Hacked–Recommends Users Set Up 2FA
And this is why we should all adopt webuathn, and get rid of totp based 2fa.

This attack vector is significantly harder to pull off if a hardware authenticator will assert that the user is logging into the correct domain.

jhgg··on Show HN: SQRL – Anti-spam rules language Twitter acquired in 2018, open-sourced
I open sourced a very minimal proof of concept for this here: https://github.com/jhgg/hyrule which shows the syntax.
← PreviousPage 3 of 17Next →