Comments URL: https://news.ycombinator.com/item?id=48422375

Points: 6

# Comments: 0

Google/Microsoft and hosting your own email is a byproduct of how difficult (socially, not technically) hosting your own email has become - mostly because SMTP protocol is inherently broken by spam and patched by social construct (trusted nodes, abuse@, 3+ DNS entries and counting, etc). Purely technical solutions, such HashCash etc, got discontinued in exchange for social ones. Central providers made (sometimes in exchange for, sometimes as excuse of, spam protection) self-hosting socially hard.

Now, I wonder if, and how, once Anthropic and OpenAI need to demonstrate profitability, could hamstring local AI. Which has been /so far/ very valuable for me in doing things that hosted providers don't want liability for, and align against (even if totally lawful and fair use!).

I wish we have agreed on some fp16/fp8 format by now to add to it..

Overall I think it’s a good thing, but it’s the closest to Eternal September I ever experienced in my life, the sudden change inclusion of cultural strangers in a shared space with its customs etc.

That said, this isn't too far from mechanical HDD latencies of the /real/ SCSI drives.

Grab a buffer, make sure it's at least as long as BHS, shove into BHS, ok, make sure it's at least as long as what the BHS requires...

Pattern matching is the hidden power of Elixir that's probably the hardest to get fluent with, it's IMHO harder than Rust's equivalent (because of gradual typing!) - but when it works it blows Pythons' match syntax way, way out.

  # Example: Login Request with immediate = 0x03 | 0x40 = 0x43.
  defp parse_lengths(<<
         _reserved::1, _i_bit::1, opcode::6,
         _flags::8,
         _specific1::16,
         ahs_length::8,
         data_length::24,
         _rest::binary
       >>) do
    {:ok, opcode, ahs_length * 4, data_length}
  end

  defp parse_lengths(_), do: {:error, :invalid_bhs}

We run "Ordered" - without queueing - so we essentially are trading off some /more/ performance for reliability in this situation. The block layer tends to handle it well in most OS.. I'll be game to test what softraid crypto would do to it (luks actually handles it reasonably!)

In terms of encrypted transport - the value is mostly in HMAC layer. That might die oddly if the ISPs decide to mess around with these packets because they look too cute - but for encrypted FS, it shouldn't be accidental, only malicious. (Accidentally, I've seen ISPs get upset over MPEG streams over netcat!)

We do suport VPD 0x83 and advertise consistent NAA/WWID, so linux will support multiple iscsiadm sessions to the same device, and it will be stitched across sessions as paths to the same disk.

We currently hardcode MC/S to 1 as part of login negotiations, advertise single portal and dropping a path will require a re-login.

So - theoretically yes, you can support multipath and it won't fall on its face, but without any practical benefits of it (no bandwidth aggregation and no failover - no ALUA) - at this point it's a single boring target.

But the underlying plumbing can support it - if you have a real usecase for it.

You can demo it out on scsipub without having to provision local SAN, and I came up with scsipub idea as a consequence of lack of any public iSCSI hosts to try things out.

Yes, it's ridiculous, but I have plugged my M1 Macbook Air to (LAN based) iSCSI through a Pi4. Works remarkably well.

Since I built it, I've started seeing it as a hammer for many nail-like problems - I think that would die down over time;

but.. I have my ESP32 "pendrive" that's net-synced. I have used it to install OS through UEFI-built-in initiator. I have added iSCSI targets to my windows laptop machine (and VMs) - while you need to deal with disconnects and reconnects, it actually works well enough.

It is a terrible idea, that doesn't sound as terrible for odd use-cases. But yes, the ESP32 over 2.4GHz over 3G internet is slow as molasses (20-30kB/s) - but when the alternative is 0.. or walking over there with a laptop, it works OK.

The other one is the Raspberry Pi{3,4,5} iSCSI shim linked there as well - I have a bunch of them for a bunch of paying clients CI/CD kinds of work, and I wanted these to boot from network, not from microSD.

Both of these projects could've benefited from a public demo iSCSI endpoint, we have http://example.com and whateveryouwant@mailinator.com - why not iSCSI

The short version: it's iSCSI targets on the public internet. Pick an image, get a block device. The free tier doesn't need a signup at all - iscsiadm -m discovery -t sendtargets -p scsipub.com and --login to iqn.2025-01.pub.scsipub:blank lands you a 64 MB scratch disk. There's a small catalog of OS images you can mount the same way.

The paid tier is where it gets less hobby-shaped: sessions survive disconnects, a single target can expose multiple LUNs, and SCSI-3 Persistent Reservations work end-to-end (REGISTER / RESERVE / RELEASE round-trip clean against sg_persist). That last bit is the cluster-storage primitive — Pacemaker, ESXi HA, and Windows MSCS all use PR for fencing — so you can actually back a 2-node failover cluster off a target on the public internet.

The post linked in the submission is the architectural decision log: Ranch 2.x listeners, a BEAM process per session, COW overlays with per-sector bitmaps, Caddy-managed Let's Encrypt for the iSCSI-TLS port without restarting the listener, and the four open-iscsi quirks that each cost me few hours. There's a section on what we're deliberately not solving (multi-region, RDMA, etc.) so you know the scope.

Two companion projects ship as embedded sub-sites on the front page — one turns an ESP32-S3 into a wireless iSCSI-to-USB bridge, one lets a Raspberry Pi 3/4/5 netboot directly from a target. Both linked from the landing page under "Hardware initiators".

Happy to answer any questions about the protocol, the deployment, or the BEAM-side design choices.

Comments URL: https://news.ycombinator.com/item?id=47948188

Points: 44

# Comments: 28

The industry is broken. It's broken in the same sense the railroad industry is broken. It has reached the point of abundance, where we're doing things that don't need doing. That won't get done in an efficient market. But since we're not in an efficient market, there are globs of capital thrown at people building stuff that.. doesn't stand a chance of actually making any return on capital.

But while it lasts, us, the glorified machine-minders (just like railroad engineers, well, minded the engines), get paid large lumps of money, through large hordes of managers, arguing on minutia of conversion optimization, and fundamentally, being paid enough to not to try and do something else, perhaps competitive.

And that is broken. Especially for the "smarter of us" - the graduation ceremony of my physics department rings true - we've trained you to discover the secrets of universe and reach the stars, and most of us will use it.. to gain an edge at Lehman Brothers.

(And I think the root of this problem, is the abundance of low-risk capital, from people who expect a small return and a pension that lasts for decades in retirement)

It’s well established that belligerents can use mines, to separate the tactical decision of deploying for purposes of area denial; from the snap-second lethal decision (if one can stretch that definition) to detonate in response to an triggering event.

Dario’s model prohibits using AI to decide between enemy combatant and an innocent civilian (even if the AI is bad at it, it is better than just detonating anyways); Sam’s model inherits the notion that the „responsible human” is one that decided to mine that bridge; and AI can make the kill decision.

How is that fundamentally different in the future war where an officer might make a decision to send a bunch of drones up; but the drones themselves take on the lethal choice of enemy/ally/no-combatant engagement without any human in the loop? ELI5 why we can’t view these as smarter mines?

The former, grants the Congress the ability to change the definition of all „lawful use” as democratically mandated (if the war is officially declared, if the martial law is officially declared).

The latter, is subtle. There can exist a human responsibility for lethal actions taken by fully autonomous AI - the individual who deploys it, for instance, can be made responsible for the consequences even if each individual „pulling of a trigger” has no human in the loop (Dario’s PoV unacceptable).

Similarly, and less subtly, acceptance of foreign mass surveillance, domestic surveillance (as long as its lawful and not meeting the unlawful mass surveillance limits!) seems to be more in the Pentagon’s favor.

Whether we like it or not, we’re heading into some very unstable time. Anthropic wanted to anchor its performance to stable (maybe stale) social norms, Pentagon wanted to rely on AI provider even as we change those norms.