Sign in

Barry Kelly

@barrkel.bsky.social
133 followers 138 following 1K posts

Software engineer and architect. Zurich, London, Galway. Google, Duco, Delphi.

PostsRepliesMedia
Barry Kelly @barrkel.bsky.social · 20h
This is in fact the clean room approach that has held up in court. However, it's normally in the form of an organization with clear division of labor and information keeping standards that can account for information flow and substantiate refutations of incorrect flow.
130
Barry Kelly @barrkel.bsky.social · 20h
Does it help that I've been dreaming about AI workflows
100
Barry Kelly @barrkel.bsky.social · 02/10/2026
You reckon? I'm just a reply guy, my dude.
100
Barry Kelly @barrkel.bsky.social · 02/10/2026
They can write programs. Configure your computer. I wanted to install GLM 5.3 flash and optimize inference for my unusual workstation. I asked Claude, went and had a shower, and came back to 35 tok/sec.
030
Barry Kelly @barrkel.bsky.social · 02/10/2026
You can stick a handful of tools with schema and docs etc in there but you rapidly run into context bloat. Better to provide an 'exec' primitive and some index, guidance and self discovery on demand. Even make it look and feel like POSIX shell, since they've been trained so hard on it.
100
Barry Kelly @barrkel.bsky.social · 02/10/2026
Shades of Memento tbh
Don't believe his lies, from Momento
040
Barry Kelly @barrkel.bsky.social · 30/09/2026
Trivial software is hyper competitive. Ambitious software is now within reach of far smaller teams, but it requires excellent product sense and risks running ahead of user feedback. I expect cheap clones of expensive complex apps that have already done the product work.
001
Barry Kelly @barrkel.bsky.social · 29/09/2026
Later Far Cry games knew they were stealth games, but they leaned too far in. Instead of emergent stealth, they were stealth minigames in a shooter shape.
001
Barry Kelly @barrkel.bsky.social · 29/09/2026
Crysis was a poor game but an excellent engine. I never replayed it at the time. By contrast, Thief was an excellent game on a poor engine. Far Cry (first half) is the best game Crytek made; Far Cry 2 is the best in the series, and it's a stealth game. You sneak past the checkpoints, dummies!
100
Barry Kelly @barrkel.bsky.social · 28/09/2026
You didn't need college for any of that though, that stuff is secondary school material. Except possibly for more interesting stats beyond binomial / Pascal's triangle and normal distribution.
100
Barry Kelly @barrkel.bsky.social · 28/09/2026
This is a pattern I think we've been seeing over and over. Big model, small model that distills what the big model has learned, more efficiently, and using user data (positive/negative) to further fine tune. Smaller model probably knows less niche stuff though.
010
Barry Kelly @barrkel.bsky.social · 28/09/2026
Fwiw mainly make these comments to find out who to mute :)
100
Barry Kelly @barrkel.bsky.social · 28/09/2026
If it helps, I don't particularly care about carbon emissions either. I don't think humanity has the collective agency to inhibit consumption enough to make a difference. I believe we'll need to come up with positive technical solutions, and not negative solutions that fail to game theory.
100
Barry Kelly @barrkel.bsky.social · 28/09/2026
I don't accept your premise; neither raising nor eating cattle is destroying the planet, nor is affecting the planet enough of a reason not not do things. We are not here merely so we may not affect anything else; if it were, suicide is the choice. The planet is doomed to stellar engulfment anyway!
100
Barry Kelly @barrkel.bsky.social · 27/09/2026
Well I do! About 5k CHF a year. Since I have zero interest in gardenjng it keeps the neighbours happy.
120
Barry Kelly @barrkel.bsky.social · 27/09/2026
Those kinds of things just stick to your teeth, they're pretty much inedible.
000
Barry Kelly @barrkel.bsky.social · 27/09/2026
Hard to be surprised when you consider people pay gardeners?
110
Barry Kelly @barrkel.bsky.social · 27/09/2026
Beef is much tastier than feed though and we don't eat for energy, we eat for pleasure!
230
Barry Kelly @barrkel.bsky.social · 26/09/2026
Everything else being equal I'd put the bed behind the chair but wouldn't forego the window when working, which demands the desk be in the middle of the room (if I read your implied window view right).
000
Barry Kelly @barrkel.bsky.social · 26/09/2026
I think it makes sense to optimize for the home office since that's likely the most frequent use (right?) and the position so you can look out the window, and not have backlight sun shining on your screen, makes sense.
100
Barry Kelly @barrkel.bsky.social · 26/09/2026
It's rather that we had 20th century thinkers who already thought through a lot of this stuff via the medium of language, and LLMs are the ultimate creatures of language. we're not starting from scratch in this.
100
Barry Kelly @barrkel.bsky.social · 26/09/2026
Those guys weren't talking about symbolic AI. Rather, Stiegler (for example) was interested in progressive exteriorization and technical reorganization of memory, language and cognition. That is acutely relevant for programmers like me who offload more to the machine.
130
Barry Kelly @barrkel.bsky.social · 26/09/2026
Derrida and Heidegger are already very relevant to AI, but especially interesting is Bernard Stiegler (unfortunate that he decided to leave just before chatgpt).
141
Barry Kelly @barrkel.bsky.social · 24/09/2026
Whatever I goddamn like
000
Barry Kelly @barrkel.bsky.social · 24/09/2026
Heh
100
Barry Kelly @barrkel.bsky.social · 24/09/2026
Wow those things have gone up in price since I bought mine last September
100
Barry Kelly @barrkel.bsky.social · 24/09/2026
You need a word that means "has a model of that attempts to predict something true in the world" that isn't "belief" or "understanding". Of course, coders have been using belief and understanding relating to entirely inanimate programs for a long time, so calling AI does not imply animism for us
000
Barry Kelly @barrkel.bsky.social · 24/09/2026
The machine is capabiltizing? The machine needs to capabilitize for several hours before producing a plan. The machine's capabilitization of the problem space is incomplete, we need to provide more details. It's not really the right word, is it?
010
Barry Kelly @barrkel.bsky.social · 24/09/2026
You're going to need words for roses that smell very similar. What do you call reasoning in text? What is it when it's written in real time? What's the correspondence between reality and the model the machine has in its computations? If you don't like the existing words, what's your alternatives?
000
Barry Kelly @barrkel.bsky.social · 23/09/2026
I personally am not hoping AI can bail us out. It's rather that AI, properly harnessed, is much more productive than hand coding, and if you don't switch you'll be left behind, a DOS programmer on Windows, a Windows programmer on the web, a desktop developer on mobile, etc.
000
Barry Kelly @barrkel.bsky.social · 23/09/2026
When asking questions about a big blob of text, I put the question at the start, paste, then repeat the question at the end. Similar intuition but also aware of attenuation of the initial question from context bloat. Also, I consistently quote big walls of pasted text with """
010
Barry Kelly @barrkel.bsky.social · 21/09/2026
Echoes of the TikTok PMs In some ways we're all PMs now www.tiktok.com/@durbinmalon...
tiktok.com
Every one of my coworkers watching this I hope you’re happy
TikTok video by Darby
000
Barry Kelly @barrkel.bsky.social · 21/09/2026
The answer to using AI right is sometimes "use even more AI". Move up the stack. Leverage more of the global stock of knowledge it has marinated in. Not asking the AI to do the whole thing from a high level spec necessarily, but learning about good leverage ideas via discussion and planning.
010
Barry Kelly @barrkel.bsky.social · 21/09/2026
Yes. It is a habit (more than a skill IMO) for SWEs to automate their jobs and it is a bit alien. Execs do a somewhat similar thing at the org level. You have to have the habit to ask for the delegation. But what is so weird about AI is you can discuss and get it to suggest delegation!
120
Barry Kelly @barrkel.bsky.social · 21/09/2026
It's tricky though. If you're doing the work yourself you can course correct and follow gradients towards interesting things. It's harder to imbue that value function into a fire and forget batch job. You can scale up, for sure, but scale too far and you risk losing control over what produces value.
130
Barry Kelly @barrkel.bsky.social · 21/09/2026
None of the others proposed to replace intellectual labor like the industrial revolution replaces physical labor. AI can't do everything yet and one of the biggest questions is if we need to solve learning efficiency, but if we have an oracle or a good simulator we can train to superhuman easily...
000
Barry Kelly @barrkel.bsky.social · 21/09/2026
I don't know how much AI you use I'm your day job, but if you used it like I do, you'd know it's no scam.
010
Barry Kelly @barrkel.bsky.social · 20/09/2026
Sometimes novelty is not actually desirable. AI usually reproduces adgacent idioms, when not given a strong steer. Can make everything seem familiar and where you expect it!
010
Barry Kelly @barrkel.bsky.social · 20/09/2026
Listen to the mathematicians. AI is not spinning fairy tales. It is doing serious superhuman work.
001
Barry Kelly @barrkel.bsky.social · 20/09/2026
AI is not a scam, Anthropic is profitable on annualized revenue, inference is a money printing machine (~90% marginal profit), and if they "pace the frontier" they cut training costs. I'm not on board with transhumanism but AI is here, not going away, and we need to deal with the change it brings.
100
Barry Kelly @barrkel.bsky.social · 20/09/2026
AI is also great at doing things you don't know how to do, because it will pick the middle of the distribution and not lead you off into the weeds. Where it'll go wrong is when your problem domain is no longeriddle of distribution and now you need something novel.
000
Barry Kelly @barrkel.bsky.social · 20/09/2026
The AI we have today can write software autonomously and we can use it to write software to generate training data for any application - software or physical hardware or biology or anything you like - that we can define a fitness function attached to ground truth for. That pie ain't air!
100
Barry Kelly @barrkel.bsky.social · 20/09/2026
How good is the AI pie? That, I think, is a big variable in how people see the world these days. If you don't like the pie, you will look for evidence of the overhang. I believe the pie will grow for quote some time, and when it slows, there will be even more people eating more of it.
100
Barry Kelly @barrkel.bsky.social · 20/09/2026
When we have something as revolutionary as AI there's a rush to capture a slice of the pie, even as the size of the pie increases. If the rate of growth of the pie declines but there's an overhang in bidding for slices, then you get a bubble pop. The pie still tastes good though.
100
Barry Kelly @barrkel.bsky.social · 20/09/2026
Note for clarity: I personally wouldn't (and don't) invest in compute hardware supply chain, because they're on the first derivative of value creation and go negative on slowdown.
110
Barry Kelly @barrkel.bsky.social · 20/09/2026
It depends on the shapes of curves: the price and cost of inference over time; and elasticity of demand for inference. It's plausible that we don't have a bubble burst dynamic at all. The strongest argument would be that latent demand for AI is elastic (superlinear response to lower prices).
100
Barry Kelly @barrkel.bsky.social · 20/09/2026
Good information about AI is not evenly distributed. There are many people who have not updated based on the advancements in the past 9 months.
100
Barry Kelly @barrkel.bsky.social · 20/09/2026
Not a great article. Newer GPUs are not actually making old GPUs obsolete these days. Not only is supply constrained, but the value that can be extracted from old GPUs is far higher than marginal cost of electricity. And performance per watt for consumer GPUs (4090 to 5090) is actually flat!
100
Barry Kelly @barrkel.bsky.social · 20/09/2026
It's not true. There is such demand that not a lot of old hardware is getting retired. In a world without supply constraints it would definitely be retired because of energy efficiency, but we're not in that world today.
210
Barry Kelly @barrkel.bsky.social · 20/09/2026
This isn't particularly true. A100 prices were $2.20/hour in September 2021 and $1.99/hour in April 2026. That's not shaped like 3 year depreciation.
130