Al Sutton

CTO at Snapp Automotive. ex-Google, ex-Facebook. The opinions stated here are my own. #tfr

Blog: https://www.alsutton.blog
GitHub: https://github.com/alsutton

Public Key

npub1jdqvsusxwz52vxprusmzsu8qpsrq7q3ya3rxl7zdff5q5kkrq5sqfx60zy

NIP-05 Address

alsutton_at_snapp.social@momostr.pink

Profile Code

nprofile1qqsfxsxgwgr8p29xrq37gd3gwrsqcps0qgjwc3n0lpx556q2ttps2gqpz3mhxue69uhhyetvv9ujuerpd46hxtnfduqs6amnwvaz7tmwdaejumr0dshct2yv

Publishing to

nostr.wine

relay.damus.io

nos.lol

Author Public Key

npub1jdqvsusxwz52vxprusmzsu8qpsrq7q3ya3rxl7zdff5q5kkrq5sqfx60zy

Show more details

Last Notes

2025-06-03 09:56:14 UTC

Small data point for folk who are playing with the #Mistral Nemo 12b Q4 #AI model.
We're now using it for code reviews at @npub1hnk…syy4, and an m4 mac mini is only 30-40% slower than a 22 core, 256GB Ram, RTX 4000 Ada workstation.
A suitable m4 mac mini costs about half of what the RTX 4000 ada GPU card alone costs.
Mac Mini's seem to be a decent way to get into playing with #LLM usage.

2025-06-02 15:05:03 UTC

I think I’ve answered my own question about whether things could go meaningfully viral on #Mastodon.
https://media.snapp.social/media_attachments/files/114/614/383/419/516/069/original/00cad2673d5f0271.jpeg

2025-05-30 18:24:20 UTC

- reply

I’m sure @npub12yu…k2zm would love another PR from an experienced contributor :)

2025-05-30 18:18:25 UTC

- reply

I’m playing with a few, but at the moment I’m preferring qwen coder 2.5 with 7b params at Q4. I’ve found some larger models generate more wordy responses, but qwen coder 2.5 was more focused on actual issues rather than vague concerns.
Here’s the Action we’re using; https://github.com/ErikHellman/claude-code-reviewer/pull/1

2025-05-30 15:54:57 UTC

- reply

Yup. The CI server is self hosted, and runs ollama locally with the appropriate model available.

2025-05-30 15:04:32 UTC

Today I deployed an #AI backed code review system as part of the @npub1hnk…syy4 #monorepo PR review checks.
Is it going to replace coders? No. Is it throwing up amazing insights? No.
Is it going to catch small things before PRs get opened for human review? Probably.... and that's where I see value. The more things an automated system catches, the fewer things humans have to spend time looking at.
#Dev #DevEx #Coding

2025-05-30 08:15:35 UTC

- reply

“What if we are the monster?” was such a twist from Rosy. 😮

2025-05-05 06:25:31 UTC

- reply

Good advice, I’d expect anyone wanting truly private conversations to use a system with E2E encryption, hence my question earlier in the thread.

2025-05-04 09:32:20 UTC

- reply

Bit of a repurposing solution, and I know you’ve had a rough time without it recently, but what about Mastodon private, targeted, messages?

2025-05-04 09:03:59 UTC

- reply

Cool. In terms of IM, my main takeaway from being the TL on the Facebook Secret Messages team many years ago, was that browsers are a pretty insecure platform if you’re trying to keep message histories secret due to the access some browser plugins have.
Do you want E2E encryption, or are you happy without it?

2025-05-04 08:49:15 UTC

- reply

I find https://european-alternatives.eu/categories useful for suggestions, and then trying things out to see what works best for me.

2025-05-03 06:20:54 UTC

- reply

https://en.m.wikipedia.org/wiki/Trooping_the_Colour

2025-05-02 16:23:29 UTC

If you vaguely thought the #UK government understood tech… https://uk.lemmy.zip

2025-05-01 05:54:55 UTC

There are many folk talking about today being the hottest start to May for the #UK since records began.
It's also worth remembering that "Net Zero" doesn't been back to previous levels, it just means we're not making things worse.
In the #UK the "Net Zero" target is 2050.
That's 25 years of things still getting worse.
25 years.
😬
https://www.bbc.co.uk/news/live/cwy0mwvx04lt
#Climate #ClimateChange

2025-04-30 08:30:40 UTC

- reply

3 is the best approach I’ve seen used anywhere. I think your choice of trade-offs is good.

2025-04-28 14:18:46 UTC

- reply

If anyone else encounters this, it’s because you have the wrong IP address for the server being given to your client 🤦‍♂️ #Debian12 #Debian #OpenSSL

2025-04-28 13:31:45 UTC

Today keeps getting stranger... next on the list, #Google pushing to the #AOSP some repo manifests which require Google Internal SSO authentication;
https://android.googlesource.com/platform/manifest/+/refs/tags/ub-automotive-master-20250418/aosp-master-with-modules.xml

2025-04-28 12:15:38 UTC

Oh this has unexpected fun written all over it....
alsutton@t5810:~$ openssl s_client -connect downloads.snappautomotive.com:443 -servername downloads.snappautomotive.com
CONNECTED(00000003)
401729AACC7F0000:error:0A000438:SSL routines:ssl3_read_bytes:tlsv1 alert internal error:../ssl/record/rec_layer_s3.c:1605:SSL alert number 80
#Debian12 #Debian #OpenSSL

2025-04-28 11:58:13 UTC

A couple of weeks after sorting out the quota issue on #Ionos (50GB hard quota on a 250GB package), but them manually increasing the quota to be, you know, actually 250GB, guess what's happened?
Yup. #Ionos's system has reverted the quota back to 50GB and I'm seeing failed upload messages again.
Time to switch to @npub18s7…wqln

2025-04-25 11:03:57 UTC

- reply

I may get some time to write a GitHub/Forgejo action which we can run on a daily basis and to get regular update PRs.

2025-04-25 08:03:02 UTC

Thanks to @npub1ang…ahch's #gradle version catalogue update plugin, I now have a replacement for #GitHub #Dependatbot, which means we're on feature parity for our #Forgejo instance 🕺 .
Thanks Hugo 🤗
https://github.com/littlerobots/version-catalog-update-plugin

2025-04-12 11:10:57 UTC

- reply

Whats the EU/UK alternative to AWS’s range of services?

2025-04-08 08:54:57 UTC

- reply

on-prem prices can go all over the place. I use workstation cards for local LLMs, and the price of an RTX 4000 (Turing generation) is pretty low.
If you're thinking of buying for on-prem then evaluating Ada Generation cards are worth it due to native support for more formats. The only down side is the stupid-ass power connector on anything higher than a RTX 2000 Ada.

2025-04-08 08:52:19 UTC

- reply

A 4b parameter model at q4 is a pretty small model (~3.3GB), so you're not really pushing limits on either.
At that level you might be seeing some accuracy issues that will need manual intervention, so going Q8 before going for higher parameter counts could be useful.
If you want to see where things differ then the 12b model will start to show things up. The CR instance could even handle 27b @ Q4, which I would expect the mac to struggle with.

2025-04-08 08:23:14 UTC

- reply

To me the test is "Is it generating faster than I can make use of the information (e.g. read it)?".
If you want to speed things up try an 8 bit quantized model (Q8), so that the weights are all 8 bit ints. The Cloud Run GPUs are Nvidia L4's, so if you're using the full precision (FP32, i.e. 32 bit floating point weights), you can speed things up by going to FP16 or even FP8.
The other thing with CR is making sure the model fits into the GPU RAM and doesn't overflow into system RAM.

2025-04-03 16:40:02 UTC

TIL: There is a category of goods in the #UK called "80% Turkeys" :O
https://assets.publishing.service.gov.uk/media/67ee668c9eae202448299c5b/request-for-input-indicative-product-list.pdf

2025-04-03 15:47:56 UTC

Moving our git repositories off #GitHub to #Forgejo is proving interesting.
The migration process Forgejo has built-in is really good, but, unfortunately, action runners have a few peculiarities which means a bit of manual fettling is needed.

2025-03-31 13:35:06 UTC

- reply

From my experience it seems a lot of the mainstream folk have moved on to BlueSky, so it tends to be more tech-knowledgeable folk here.
If tech-knowledgeable fits the community, there's a number of fediverse compatible implementations listed at https://www.reddit.com/r/fediverse/comments/1i9bgs8/list_of_fediverse_software/ that focus on different types of media.

2025-03-27 22:06:01 UTC

- reply

Is there a guide others can follow that shows how to keep issues, PRs, etc.?

2025-03-27 09:02:46 UTC

- reply

#define OK 200
what could possibly go wrong? :)

2025-03-27 07:29:12 UTC

Nothing annoys me more that pop-up, modal, adverts in a service that is already being paid for.
#Google #Workspace #NoToGemini
https://media.snapp.social/media_attachments/files/114/233/206/809/187/956/original/099a23b9b1d3fa4e.png
https://media.snapp.social/media_attachments/files/114/233/216/206/342/986/original/ccbf55c4d391ad5e.png

2025-03-26 19:42:56 UTC

The next step in the great, long-term, #Android bait & switch :( https://www.androidauthority.com/google-android-development-aosp-3538503/

2024-12-31 18:57:17 UTC

- reply

Completely agree, but I’m guessing Apple needs a way to stop people signing up, binge watching in a month, then cancelling.

2024-12-31 18:54:22 UTC

- reply

Silo is worth a try.

2024-11-21 19:48:00 UTC

Looks like the Khadas Vim3 board is on sale in a popular global marketplace .
That board seems to run #Android well and is on the #AOSP reference board list.
Personally I’d pick the “Pro” version. 😉
https://source.android.com/docs/setup/create/devices

2024-09-09 07:51:16 UTC

- reply

I never thought we'd reach a point where cars lose software features as they get older :(

2024-09-08 15:50:12 UTC

… and so the sun sets on my #UK holiday, and it’s back to work tomorrow.
If you’re not had a holiday in a while, then just do it. If nothing else it helps to identify things that your on the hot path for, which you should be looking to have a “back up” option for.
https://s3-eu-west-2.amazonaws.com/snapp-social-files/media_attachments/files/113/102/723/986/471/868/original/8f2d40c5c93c93fc.jpeg