Brave thing to own up here. Hope you don't get too much flack. I suspect no-one will be thinking harder about LLM summaries and journalism in future articles, than you!
Meanwhile I feel like we should all do that "HBO intern sends a test email" thing and send times we've screwed up with LLMs!
I think apophis standing on the balcony of a national trust place could be just the look their series needs
๐ welcome to GDS
The bit I've never figured out is, where do all these horses skulls come from? More questions follow every answer with Mari Lwyd
Very excited to hear @nicolefv.bsky.social and Abi Noda have a new book. Accelerate is already a firm favourite.
Wonder if there are plans for an audiobook like previous books? I've enjoyed flipping between the paperback as reference and audio when out and about.
It certainly is useful, I've been using that old presentation work to unpack how it all works ever since you made it.
Very useful to explain how this all works to new recruits too. Very tricky to decode from the outside I think. I do check for updates but seemed all good last I checked.
Personal data storage is an idea whose time has come: Data Ownership as a conversation changes when data resides primarily with people-governed institutions rather than corporations.
https://blog.muni.town/personal-data-storage-idea/
Now I've asked this, slightly embarrassed to find the answer is "yes, Huw, your new department makes one"
You given it a go? Does it feel useful / fun?
Well, until it starts making up transcripts, or giving me dreamed up renders of PMQs...
@alexparsons.bsky.social ๐-ing recent MySociety work referencing AI, you know of any attempts to fine tune an LLM across the whole of Hansard?
Having a chat with a bot able to reference the history of a debate or topic feels like one of the few LLM use cases I'd be super interested in playing with
But then,
Headlines below about "good jobs replaced by AI" as image detection of Matt Damon now comes as standard in all CCTV.
Anywhere he walks the clip bellows from tinny speakers
This thread reminds me a lot of this take from the height of the RTO mandate days:
www.makeworkbetter.info/p/fixing-cul...
Particularly the stat we're in 20% more meetings than pre-pandemic (I think according to an MS study?).
I do think we need to ask more "what are meetings for" in 2025
Amid all the UK chat about steel this week I went back to listen to the second half of foreignpolicy.com/podcasts/one...
Feel like it's the primer for global steel industry I've been missing in other reports. Tune in about 15 mins in after the Davos segment
I also recall hearing of a theoretical attack where you like... Put a novel in a field for something that doesn't truncate... And then things fall over in the processing, but that feels a bit unlikely now I say it out loud... Unless anyone has a better insight?
But... There are plenty of other algos around now, I don't know how much this is an old problem or a present one.
But if you use funky characters you could go over 72 bytes with something like 36 chars. Or 18 emojis.
So unicode and pw hashes might give some cause to upper limits, sometimes, no?
Back when I used to do more Rails, you could end up with a Bcrypt truncation problem. The maximum size of a Bcrypt hash is 72 bytes so it truncates and ignores anything longer. False sense of security for folks with passwords over that (it's all just 72 bytes), perhaps surprise behaviour etc.
Feels like an opportunity there for business that don't need the surveillance, or government websites to showcase what's possible IMO
Really good, cheers for the link
GDS is hiring a senior content designer to work on the Service Standard and Service Manual.
ยฃ56,070 - ยฃ70,219
Deadline end of Wednesday.
www.civilservicejobs.service.gov.uk/csr/jobs.cgi...
#govdesign
So perhaps the challenge here is, not are we OK with granting tool access to see what the vibes can bring.
But does doing that, for folks with less appreciation of the risks of some of these tools mean you give up on the idea of managed devices and try to zero trust everything. Vibe code included?
Yeah I could get behind this, following the "disposable spreadsheet" model. I think it'd be a learning curve but probably good for future skills.
I don't think organisations are ready for that though. Like a spreadsheet doesn't have a package ecosystem that lets you install and run anything.
Usual caveats being, many things in spreadsheets should be in products / databases etc etc. sure folks will make a mess too.
But I'd wanna see benefits realised in excel functions before we start asking if LLMs means everyone should get a python env
The single biggest productivity win I've had from LLMs is getting them to help write Google Sheets and Excel queries for disposable, quick fix spreadsheets. If the pitch was AI will democratise data wrangling in sheets. I'd be in, and say it should be broadly accessible.
Plenty of us are willing to share the dark arts if you want to join the priesthood
Getting properly to grips with ActivityPub is on my 2025 list.
What's the best onramp project or minimums viable setup to start playing right now?
We should ensure this is on #openbenches openbenches.org
Currently trying the Murderbot diaries, which are fun and short too
Things I've read in the last few years I gave 5* to.
- Adrian Tchaikovsky's children of time/ruin/memory series.
- Re-read a load of Ian M Banks culture books, still great.
- The expanse series.
- Ann Leckie's Ancillary series is fun too.
Pick any random one and it's a fun diversion I think
I would be interested to know from folks more into AI how this generally works in terms of compute demand? I imagine primary model training is hungry... but is that really what folks are doing with all those Nvidia chips?
Or is that a one off and then it's all burning cycles for fine tuning ๐ค
