HNHacker News
TopNewBestAskShowJobs

benlivengood

2,633 karma · joined August 13, 2019

submissionscomments
benlivengood··on Mind the encryptionroot: How to save your data when ZFS loses its mind
I can only think of one situation where it's hard to retrieve a remote backup of an individual file; where ZFS native encryption is in use and the remote backup system is not trusted to load the key for the dataset.

For myself, I don't trust remote systems to always have keys loaded, but in an emergency I would feel relatively safe temporarily loading the key, mounting the snapshot read-only, and scp-ing the files out, then unmounting and unloading the key (and rebooting for good measure).

There's also a viable slow option; export the raw storage of the backup ZFS pool over the (inter)network to a trusted machine and import the pool read-only locally, load the key, mount the filesystem, and make a copy. Much slower but is practical. I've used s3backer fairly successfully as a backup method for a pool with native encryption; it takes a minute or so to import the pool and can write backup snapshots at a few MB/s, so there shouldn't be any fundamental reason iscsi or similar wouldn't work.

benlivengood··on The Beginner's Textbook for Fully Homomorphic Encryption
The difference between homomorphic schemes and fully homomorphic schemes is that FHE can be bootstrapped; there's a circuit that can be homomorphically evaluated that removes the noise from an encrypted value, allowing any homomorphic calculation's result to have its noise removed for further computation.
benlivengood··on Meta’s live demo fails; “AI” recording plays before the actor takes the steps
CEOs (and other spokespeople) are actors paid to believe things convincingly in front of other people.
benlivengood··on Novel hollow-core optical fiber transmits data faster with record low loss
I think you could moderate a large nuclear fission reactor (on the edge of criticality) to produce detectable differences in neutrino rate on the other side of the world. KamLAND and a few other experiments detected multi-GW reactor anti neutrinos at 1000km[0] and so presumably tens of GW (easily [if perhaps not safely?] achievable briefly in current reactors) should be detectable over 8000km.

[0] https://www.sciencedirect.com/science/article/pii/S173857331...

benlivengood··on Google deletes net-zero pledge from sustainability website
I'd wager the biological cellular economy is more efficient and more robust than capitalist economies.

Not that I advocate nature red in tooth and claw as the epitome of governance.

benlivengood··on The Bitter Lesson Is Misunderstood
I don't think anyone has yet trained on all videos on the Internet. Plenty of petabytes left there to pretrain on, and likely just as useful once the text/audio/image pretraining is done.
benlivengood··on Abstract Machine Models (2022)
I think a formalism for AMMs might go a long way toward making formal software development more straightforward. All the formalisms I am aware of pair some logic with some programming language, or worse a logic with an abstract model of a language, but implicitly or explicitly portions of architecture get dragged into the formalism as well, e.g. resulting in seL4 having differing levels of proofs for various architectures (arm, amd64, risc-v).

If instead we pair a logic with a parameterized (by cache coherency semantics, hardware protection and isolation levels, register layout, etc.) abstract machine model then it could presumably apply over multiple architectures and languages with derived specializations for each language.

benlivengood··on It is worth it to buy the fast CPU
It's straightforward to measure this; start a stopwatch every time your flow gets interrupted by waiting for compilation or your laptop is swapping to keep the IDE and browser running, and stop it once you reach flow state again.

We managed to just estimate the lost time and management (in a small startup) was happy to give the most affected developers (about 1/3) 48GB or 64GB MacBooks instead of the default 16GB.

At $100/hr minimum (assuming lost work doesn't block anyone else) it doesn't take long for the upgrades to pay off. The most affected devs were waiting an hour a day sometimes.

This applies to CI/CD pipelines too; it's almost always worth increasing worker CPU/RAM while the reduction in time is scaling anywhere close to linearly, especially because most workers are charged by the minute anyway.

benlivengood··on Evaluating LLMs playing text adventures
Wouldn't playthroughs for these games be potentially in the pretraining corpus for all of these models?
benlivengood··on My Lethal Trifecta talk at the Bay Area AI Security Meetup
I'm out of this game now, and it solved a very particular problem in a very particular way with the current feature set.

See sibling-ish comments for thoughts about what we need for the future.

benlivengood··on My Lethal Trifecta talk at the Bay Area AI Security Meetup
I believe you've covered some working solutions in your presentation. They limit LLMs to providing information/summaries and taking tightly curated actions.

There are currently no fully general solutions to data exfiltration, so things like local agents or computer use/interaction will require new solutions.

Others are also researching in this direction; https://security.googleblog.com/2025/06/mitigating-prompt-in... and https://arxiv.org/html/2506.08837v2 for example. CaMeL was a great paper, but complex.

My personal perspective is that the best we can do is build secure frameworks that LLMs can operate within, carefully controlling their inputs and interactions with untrusted third party components. There will not be inherent LLM safety precautions until we are well into superintelligence, and even those may not be applicable across agents with different levels of superintelligence. Deception/prompt injection as offense will always beat defense.

benlivengood··on My Lethal Trifecta talk at the Bay Area AI Security Meetup
Dia is currently (as of last week) not vulnerable to this kind of exfiltration in a pretty straightforward way that may still be covered by NDA.

These opinions are my own blah blah blah

benlivengood··on Make Your Own Backup System – Part 2: Forging the FreeBSD Backup Stronghold
For the receiving side the user needs `receive` permission for the dataset to receive new snapshots, but it's convenient to grant `create,mount` as well so the user can send new child datasets which may have mount options in the source dataset, relying on inheritance for the allow permissions to apply to the child datasets.

On the sending side `send` is enough, but for tools like syncoid and znapzend `hold,release` are useful as well since typically they hold the latest snapshot on the source which the destination also has so that it can't be deleted on the source before it's used to send an incremental stream up to a newer snapshot only available on the source.

benlivengood··on Does the Bitter Lesson Have Limits?
I think it's a little early (even in these AI times) to call HRM a counterexample of the bitter lesson.

I think it's quite a bit more likely for HRM to scale embarrassingly far and outstrip the tons of RLHF and distillation that's been invested in for transformers, more of a bitter lesson 2.0 than anything else.

benlivengood··on Make Your Own Backup System – Part 2: Forging the FreeBSD Backup Stronghold
I've had good luck using `zfs allow` to grant non-root backup users the ability to only add snapshots to their datasets to avoid the "attacker compromises prod and then jumps to the backup server and deleted the backups". It is an extra step to clean up old snapshots, but worth the risk-reduction.

You can also split administration up so that, e.g., my friend sending me snapshots can't even log in as root on his backup server.

benlivengood··on How did Facebook intercept competitor's encrypted mobile app traffic? (2024)
All cloud services are in a similar position; they hold the private TLS keys and could reveal them in response to legal process, allowing active MITM (perfect forward secrecy prevents passive data theft without more intrusive realtime access to VM RAM).

Only a very specific configuration of "Confidential Computing" (based on AMD SEV or Intel TDX) where boot attestation is checked remotely before private keys are sent from an on-premise store (or a fully trusted hosted HSM) to the remote VM could prevent a cloud provider from intercepting private key material, and only then as far as boot attestation and SEV/TDX is trusted.

benlivengood··on Sign in with Google in Chrome
Why would you share your real name with Google when making a gmail account, or use your real picture?

It's fine to be pseudonymous on the Internet if you are in control of your pseudonyms, which Google accounts actually does allow with some extra work (don't mix your chrome profiles and Google accounts, etc.)

Or, like me, you can roll the dice on real names on the Internet (for professional things mostly)

benlivengood··on Do not download the app, use the website
Even Signal asks for Contacts. Whatsapp asks every other time you open the app.

I can't use Zelle on my bank's web page any more, they just redirect to their app which is literally just their website in an app.

benlivengood··on Never write your own date parsing library
Unfortunately, the Second is measured for purposes of our timekeeping standards at sea-level on Earth which is ~1PPB slower than it would be in free space, as opposed to having a correction factor built into our time standards and so, for example, interplanetary ping times would be slightly shorter (in UTC/TIA nanoseconds) than expected.
benlivengood··on Never write your own date parsing library
I would not be surprised if the state of unit tests on good date parsing libraries are not sufficient to design a new one from scratch.

See the number of unit tests in the Linux kernel, for example.

benlivengood··on CO2 Battery
I'm guessing the diagram is missing a bit on the heat exchanger side; they're going to need to dump plenty of (environmental) heat into the expansion thingy to keep the liquid CO2 boiling off indefinitely at the pressure they want.

If this is intended for small-scale to medium-scale on-premise storage then the evaporating CO2 could also serve as the cold side of a building-size AC system for extra efficiency during the high demand portion of the duck curve.

I think there may be quite a market for maintaining hot and cold (and pressurized/liquified) sinks throughout the day/night cycle in highrises or entire cities.

benlivengood··on Women dating safety app 'Tea' breached, users' IDs posted to 4chan
I think iOS and Android already holds the threat of app store removal over developers' heads.

Presumably the risk/reward still favors risky practices.

benlivengood··on Women dating safety app 'Tea' breached, users' IDs posted to 4chan
In this case it appears to be a public Firebase bucket; shutting down the app wouldn't help. Quite possibly access to Firebase was mediated through a backend service and Apple couldn't validate the security of the unknown bucket anyway.
benlivengood··on Quantitative AI progress needs accurate and transparent evaluation
I think using LLMs/AI for pure mathematics is one of the very least ethically fraught use-cases. Creative works aren't being imitated, people aren't being deceived by hallucinations (literally by design; formal proof systems prevent it), from a safety perspective even a superintelligent agent that was truly limited to producing true theorems would be dramatically safer than other kinds of interactions with the world, etc.
benlivengood··on It's a DE9, not a DB9 (but we know what you mean)
And Molex power connectors are actually AMP Mate-n-Lok connectors.

I didn't learn this until this year...

benlivengood··on Two narratives about AI
If you're like me, though, your day-to-day has probably shifted a lot.

My day used to be making sure desktops worked and we had a repeatable process to make new good desktops out of all the complex client software they needed.

Then I made sure servers got upgraded and patched and taken care of on a day-to-day level, although it was still someone else's job to keep desktops running. At home I compiled my own kernels and used tarballs to install and update packages. Desktop hardware support was iffy in Linux.

Then I jumped to borg and tupperware and kubernetes where hardware never mattered and it almost didn't matter what clients had because the browser they used auto-updated. At home I switched to distros where package management was automatic and rarely, if ever, broke.

I don't even know the hostnames or the network addresses of the hardware that runs my services, and AWS or GCP SREs probably rarely need to know either. Now I care about an abstract thing called a service that is instrumented with logs, metrics, and traces that would put the best local development tools of 20 years ago to shame. CI/CD and infrastructure as code pipelines actually did automate away many of the checklist-style sysadmin work of the past. At home I could run Talos and Ceph and Crossplane if I wanted to but so far I've dragged the old days of individual hosts along mostly for nostalgia.

I expect to eventually end up caring about systems at an even more abstract level once something like Crossplane becomes as universal as Terraform and GitHub actions. They'll probably run on something like web assembly on bare metal because no one actually cares what's underneath their containers if they keep working.

The stack of technology gets taller and more abstract and as it does so the job of caring about the lower layers gets automated and the lower layers get, if not simplified, standardized to the point that automation is more reliable than human intervention.

Humans will only get squeezed out of the loop when superhuman artificial intelligence arrives and our abstract design and management of systems becomes less reliable than the automation. Then hopefully we get a nice friendly button to push for more automatically-human-aligned utility.

EDIT: That's not to say that the lower levels can be automatically designed; not yet at least. Eventually once AI is good enough at formal design then quite likely. We still need low-level software engineering to keep building the stack but it is vastly more commoditized (the one open source developer in the xkcd cartoon keeping 99% of the world's infrastructure running on a 20-year old tool/utility)

benlivengood··on Parsing Protobuf like never before
Or, you know, several hundreds of billions of times per second...
benlivengood··on Migrating a ZFS Pool from RAIDZ1 to RAIDZ2
Administrative risk, mostly. If you accidentally wipefs the wrong disk when moving it to the RAIDZ2 pool or offline a real disk instead of a tmp disk or make some other simple mistake then a zpool might be unrecoverable.

An alternative would have been to build a RAIDZ2 out of three new disks and two tmp files and then copy the data over and finally offline a RAIDZ1 disk and online it on the RAIDZ2 zpool. Two copies of the data in each pool would require an actual disk failure in both zpools to lose data during the resilvering (even though both were degraded), and when breaking the original zpool to add a replacement disk for the 5th device in the RAIDZ2 pool you'd have to lose two existing disks in that pool to lose data.

No criticism; it's cool to be able to do moves with minimal resources and I've also thought about potential ways to upgrade zpools under weird constraints; especially no free SATA ports for example. I had thought about using the four SATA ports to copy data from half the source disks in a RAIDZ2 to half the destination disks of a new RAIDZ2 while the removed disks would still be a functional zpool on their own as a backup. But ultimately I found a cheap extra box and went for a complete upgrade to larger disks with the benefit of keeping the old box as a local mirror.

My upgrade was from a single RAIDZ2 of 4x 4TB disks (8 usable) to a RAIDZ1 of 4x 14TB disks (42 usable) because once both pools existed I was comfortable running the old disks in a 12TB RAIDZ1 to receive snapshots from the primary zpool while I had <12TB used. Still requires 2 disks to fail to lose data, and now I do some offsite backups as well (rsync and s3backer) and have a test system to perform upgrades on before the main system.

benlivengood··on LetsEncrypt Outage
Hopefully the thundering herd when service is restored doesn't knock things offline again. I know LE designs for huge throughput (something like 3X total outstanding certificates in 24 hours, at one point) and the automated client recommendations for backoff are pretty good, but there will be a lot of manual applications/renewals I'm sure.
benlivengood··on In the long run, GPL code becomes irrelevant (2015)
That's why I used the scare quotes; I don't have a hard and fast rule here for what's legitimate market-making software value and what's just rent-seeking.

I think that much of the core features in Desktop Photoshop-style products or video games probably requires legitimate short copyright, for example. I don't think copyrighting a plugin interface or header files or hardware drivers makes any sense, unless there's significant creative work like e.g. the portions of GPU drivers that trade off appearance and speed in novel ways.

← PreviousPage 5 of 34Next →