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

[email protected]

Profile Code

nprofile1qqsfxsxgwgr8p29xrq37gd3gwrsqcps0qgjwc3n0lpx556q2ttps2gqpzpmhxue69uhkummnw3ezuamfdejsz9rhwden5te0wfjkccte9ejxzmt4wvhxjmceqxa0d

Publishing to

nostr.wine

relay.damus.io

nos.lol

Author Public Key

npub1jdqvsusxwz52vxprusmzsu8qpsrq7q3ya3rxl7zdff5q5kkrq5sqfx60zy

Show more details

Last Notes

2025-07-02 12:06:01 CEST

Oh the joy of having to block AI bots from scraping some public repos we host from the bot with the signature
"KHTML, like Gecko; compatible; GPTBot/1.2; +https://openai.com/gptbot"
I hope they read the license for all the code before using it ;)

2025-06-26 18:22:46 CEST

Oh great, now I can’t trust the emails from #Google not to change between reading.
One of the reasons I value email is for the ability to store what’s been said at a point in time and refer back to it knowing it will not change, kind of like, books or letters.
https://media.snapp.social/media_attachments/files/114/750/584/367/493/530/original/d64f243425245236.png

2025-06-18 12:29:14 CEST

- reply

I’ve not seen anything using it so far. I suspect there’ll be some time to see how it develops before there’s a large scale buy-in.

2025-06-13 13:04:00 CEST

Oh man, this is a trip down memory lane… https://m.youtube.com/watch?v=buuUZFh_pyk

2025-06-13 12:18:48 CEST

There are some things that you should just not get involved in. https://apple.news/A7_mxU4BeRyK22qHiCeasLg

2025-06-09 18:57:53 CEST

Having to add an explicit dependency declaration to a #gradle build file, to ensure that an app isn't built with a GSON version that has vulnerabilities dating back to 2022, feels sooo wrong..... but hey, that's #AndroidDev!
(if you're wondering, a dependency of the project has a deep transitive dependency on GSON 2.8.9, and Gradle is happily hauling that in to the build with no warnings 😬 )

2025-06-03 11:56:14 CEST

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.

2025-06-01 13:56:15 CEST

This should be on TV every day, possibly multiple times per day…. #AI
https://media.snapp.social/media_attachments/files/114/607/978/725/543/871/original/f9489651f81bdd9b.mp4

2025-05-31 17:34:44 CEST

- reply

I’ guessing the Claud hardware is high end, so i’d a lot more performant. The “num_ctx” option, which is set in the request the action makes, lets you change the context window, but you bigger windows slow down the review.

2025-05-30 20:24:20 CEST

- reply

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

2025-05-30 20:18:25 CEST

- 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 17:54:57 CEST

- reply

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

2025-05-30 17:04:32 CEST

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

2025-05-26 23:02:02 CEST

- reply

Gitea / @npub12r3…zz58 use a close compatible GHA language for CI/CD s standard, and have an action runner written in Go.

2025-05-10 12:21:30 CEST

- reply

I have a feeling the Swiss connection could be for tax reasons. The history is very US focused.
https://media.snapp.social/media_attachments/files/114/483/035/332/580/047/original/871c38755fb73cc4.jpeg

2025-05-06 08:22:00 CEST

Welcome to the work #Android15 #AAOS #LTS
https://android.googlesource.com/platform/manifest/+/refs/tags/android-automotiveos-15.0.0_lts1

2025-05-04 11:32:20 CEST

- 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 11:03:59 CEST

- 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 10:49:15 CEST

- reply

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

2025-05-02 18:23:29 CEST

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

2025-05-01 07:54:55 CEST

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 10:30:40 CEST

- reply

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

2025-04-28 15:31:45 CEST

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 14:15:38 CEST

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 13:58:13 CEST

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 13:03:57 CEST

- 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 10:03:02 CEST

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

2025-04-21 10:20:30 CEST

Given all the “leaders” of large groups of people that are around at the moment, he seemed like a good guy. https://www.bbc.co.uk/news/live/crknlnzlrzdt

2025-04-20 20:40:08 CEST

I feel sad every time I take folk through the arc of the web being a place where anyone could publish, and popularity was based on how many folk linked to your page because it was interesting, to today, when most folk consume content that’s been promoted using an algorithm that can be tweaked to the interests of a company owner or maximising advertising accenture.

2025-04-17 15:55:05 CEST

TIL: As of 14th March, you can no longer use any #POP3 or #IMAP client to access emails held in Google Workspace; https://workspaceupdates.googleblog.com/2023/09/winding-down-google-sync-and-less-secure-apps-support.html

2025-04-14 19:58:44 CEST

- reply

Yup, and our Managed NextCloud membership, which is supposed to include 3TB of storage, is currently stuck at a 15GB quota (so only 5% of the advertised amount)

2025-04-14 10:29:24 CEST

After a week, and conversations with various folk at #Ionos, I have found someone who knows what's going on.
They do, as I suspected, implement "increasing quotas", which means that initially folk are initially provisioned with low quotas, and those are adjusted, in my case, on a daily basis.
So even though the web hosting package is advertised as having 250GB available, the quote limited it to 50GB, then when we hit that increased it to 70GB.
It's now been manually set to 250GB.

2025-04-13 11:35:16 CEST

When #Amazon starts showing return information as a part of listings, you know it’s got problems.
https://media.snapp.social/media_attachments/files/114/329/970/992/381/321/original/95f0bdd14c1c51cd.png

2025-04-12 13:10:57 CEST

- reply

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

2025-04-08 10:54:57 CEST

- 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 10:52:19 CEST

- 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 10:23:14 CEST

- 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 18:40:02 CEST

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 17:47:56 CEST

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.

2025-03-27 10:02:46 CET

- reply

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

2025-03-27 08:29:12 CET

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 20:42:56 CET

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

2025-03-10 12:05:14 CET

I'm surprised at the number of people who still point to social media accounts on their slide decks rather than buying light-weight hosting for a personal domain and using that in their decks.
Recent history has shown that social media companies change, and your old decks can become dead-ends.

2024-12-31 19:57:17 CET

- reply

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

2024-12-13 20:05:18 CET

So how many months/years do we think it’ll be before #Android XR is abandoned like Tablets, Stadia, etc., etc., etc.?

2024-11-21 20:48:00 CET

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

2024-11-03 16:40:00 CET

- reply

- Real time updates, including rolling out patches quickly (no OTAs)
- 3D graphics
- Streaming large amounts of data from storage (Usually a tape at 9600 baud which the user had to wait for)

2024-09-09 09:34:14 CEST

Does anyone at #Volvo really believe that, after 4 years, folk will pay a subscription to access services you can get for free via #AndroidAuto?
I can’t see locking this functionality behind a paywall, after 4 years, playing out well with consumers.
https://s3-eu-west-2.amazonaws.com/snapp-social-files/media_attachments/files/113/106/436/183/529/168/original/52b8732126b5080d.jpeg

2024-09-08 17:50:12 CEST

… 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