Rob Bowley
@robbowley.net
Product & Tech Leadership Advisor, Consultant, Coach & Mentor
Tech, Software Development, Science, History, Economics, Politics
https://blog.robbowley.net
https://www.linkedin.com/in/robertbowley
https://pragmaticpartners.co.uk
Manchester, UK
Tech, Software Development, Science, History, Economics, Politics
https://blog.robbowley.net
https://www.linkedin.com/in/robertbowley
https://pragmaticpartners.co.uk
Manchester, UK
Pinned
Rob Bowley
@robbowley.net
· Nov 16
Lots of new followers so thought I'd do an intro 🙂
This is me, without hat, in my garden, on a rainy morning in Manchester 🐝, UK
Been working in tech for 25 years - software engineer then various leadership roles. Nowadays I'm a product & tech leadership advisor, coach, consultant.
1/4
This is me, without hat, in my garden, on a rainy morning in Manchester 🐝, UK
Been working in tech for 25 years - software engineer then various leadership roles. Nowadays I'm a product & tech leadership advisor, coach, consultant.
1/4
As dust settling on ChatGPT5.2 and Gemini 3, the pattern appears to be capability reallocation rather than general improvement.
Better scores on narrow reasoning and benchmarks, but trade offs elsewhere like consistency & hallucinations.
Better scores on narrow reasoning and benchmarks, but trade offs elsewhere like consistency & hallucinations.
December 18, 2025 at 5:07 PM
As dust settling on ChatGPT5.2 and Gemini 3, the pattern appears to be capability reallocation rather than general improvement.
Better scores on narrow reasoning and benchmarks, but trade offs elsewhere like consistency & hallucinations.
Better scores on narrow reasoning and benchmarks, but trade offs elsewhere like consistency & hallucinations.
Reposted by Rob Bowley
I see a lot of complaints about untested AI slop in pull requests. Submitting those is a dereliction of duty as a software engineer: Your job is to deliver code you have proven to work simonwillison.net/2025/Dec/18/...
Your job is to deliver code you have proven to work
In all of the debates about the value of AI-assistance in software development there’s one depressing anecdote that I keep on seeing: the junior engineer, empowered by some class of …
simonwillison.net
December 18, 2025 at 2:57 PM
I see a lot of complaints about untested AI slop in pull requests. Submitting those is a dereliction of duty as a software engineer: Your job is to deliver code you have proven to work simonwillison.net/2025/Dec/18/...
From fire and cooking, agriculture, to fossil fuels, the biggest jumps in human living standards came from unlocking more usable energy.
Any future industrial-revolution-scale shift will almost certainly involve a comparable change in energy availability, not information technology alone.
Any future industrial-revolution-scale shift will almost certainly involve a comparable change in energy availability, not information technology alone.
Faster horses, not trains | Rob Bowley
blog.robbowley.net
December 18, 2025 at 1:38 PM
From fire and cooking, agriculture, to fossil fuels, the biggest jumps in human living standards came from unlocking more usable energy.
Any future industrial-revolution-scale shift will almost certainly involve a comparable change in energy availability, not information technology alone.
Any future industrial-revolution-scale shift will almost certainly involve a comparable change in energy availability, not information technology alone.
How funny this comes up in my feed when I've just written a piece saying GenAI is not like steam or electricity
December 17, 2025 at 3:34 PM
How funny this comes up in my feed when I've just written a piece saying GenAI is not like steam or electricity
How AI Is Transforming Work at Anthropic
How AI Is Transforming Work at Anthropic
www.anthropic.com
December 17, 2025 at 1:45 PM
I’ve been trying to work out why new GenAI models don’t feel much different to me
GenAI operates through a narrow, lossy interface. Good at helping with slices of work, but real constraints sit elsewhere
Better models don't change the boundary. So it doesn’t really matter how “smart” models get
GenAI operates through a narrow, lossy interface. Good at helping with slices of work, but real constraints sit elsewhere
Better models don't change the boundary. So it doesn’t really matter how “smart” models get
Faster horses, not trains | Rob Bowley
blog.robbowley.net
December 17, 2025 at 12:47 PM
I’ve been trying to work out why new GenAI models don’t feel much different to me
GenAI operates through a narrow, lossy interface. Good at helping with slices of work, but real constraints sit elsewhere
Better models don't change the boundary. So it doesn’t really matter how “smart” models get
GenAI operates through a narrow, lossy interface. Good at helping with slices of work, but real constraints sit elsewhere
Better models don't change the boundary. So it doesn’t really matter how “smart” models get
GenAI feels like faster horses, not trains. Powerful and useful, but not yet the kind of change you can point and say “this could not exist before”
1/4
1/4
Faster horses, not trains. Yet | Rob Bowley
blog.robbowley.net
December 16, 2025 at 2:52 PM
GenAI feels like faster horses, not trains. Powerful and useful, but not yet the kind of change you can point and say “this could not exist before”
1/4
1/4
The industrial revolution broke hard limits. You could point at a train and say this could not exist before - people and goods moving in hours, not days.
GenAI = more like faster horses - better versions of what we already had: writing, code, analysis, planning
Useful, but not the kind of shift
GenAI = more like faster horses - better versions of what we already had: writing, code, analysis, planning
Useful, but not the kind of shift
December 16, 2025 at 9:36 AM
The industrial revolution broke hard limits. You could point at a train and say this could not exist before - people and goods moving in hours, not days.
GenAI = more like faster horses - better versions of what we already had: writing, code, analysis, planning
Useful, but not the kind of shift
GenAI = more like faster horses - better versions of what we already had: writing, code, analysis, planning
Useful, but not the kind of shift
Observation: I don't think AI having a habit of creating duplication in code is just a consequence of being trained on poor code. When I'm using it to support writing/documentation, it's constantly repeating and duplicating things.
December 11, 2025 at 1:53 PM
Observation: I don't think AI having a habit of creating duplication in code is just a consequence of being trained on poor code. When I'm using it to support writing/documentation, it's constantly repeating and duplicating things.
Thought's in my head today, now on paper.
AI will more likely mean more with the same rather than the same or more with less.
AI will more likely mean more with the same rather than the same or more with less.
More with less, or is it more with the same | Rob Bowley
blog.robbowley.net
December 11, 2025 at 9:48 AM
Thought's in my head today, now on paper.
AI will more likely mean more with the same rather than the same or more with less.
AI will more likely mean more with the same rather than the same or more with less.
Good write up on/experience report on working with Claude Code.
Only one thing I strongly disagree with:
"Claude is trained on real-world code. Real-world code has crappy tests."
Most real-world code has NO tests.
Only one thing I strongly disagree with:
"Claude is trained on real-world code. Real-world code has crappy tests."
Most real-world code has NO tests.
How I Built a Production App with Claude Code
The follow-up everyone asked for...
leadershiplighthouse.substack.com
December 10, 2025 at 2:16 PM
Good write up on/experience report on working with Claude Code.
Only one thing I strongly disagree with:
"Claude is trained on real-world code. Real-world code has crappy tests."
Most real-world code has NO tests.
Only one thing I strongly disagree with:
"Claude is trained on real-world code. Real-world code has crappy tests."
Most real-world code has NO tests.
One of the interesting things from picking up running this year - not how fast I can run at full effort, but I can now run 5k in ~30 mins whilst barely getting out of breath or breaking into a sweat
December 10, 2025 at 7:05 AM
One of the interesting things from picking up running this year - not how fast I can run at full effort, but I can now run 5k in ~30 mins whilst barely getting out of breath or breaking into a sweat
Reposted by Rob Bowley
🧵 As we figure out what role LLMs and Gen AI systems will play in developing software, I await with interest actual experience reports, such as this one:
(> next)
(> next)
December 9, 2025 at 2:51 PM
🧵 As we figure out what role LLMs and Gen AI systems will play in developing software, I await with interest actual experience reports, such as this one:
(> next)
(> next)
I tried this with ChatGPT. Spent a long time building up a profile of a relative who is hard to buy for. All I got were pretty generic bland suggestions, none were any good - nothing much better then e.g. Google Shopping or web search
www.bbc.co.uk/news/article...
www.bbc.co.uk/news/article...
AI tools transform Christmas shopping as people turn to chatbots
Shopper are increasingly using chatbots for their hard-to-buy-for family and friends, with implications for bargain hunters and businesses.
www.bbc.co.uk
December 9, 2025 at 3:30 PM
I tried this with ChatGPT. Spent a long time building up a profile of a relative who is hard to buy for. All I got were pretty generic bland suggestions, none were any good - nothing much better then e.g. Google Shopping or web search
www.bbc.co.uk/news/article...
www.bbc.co.uk/news/article...
Good guide if you're needing to create policy for use of LLMs at your org (which you should)
I have put together a (long overdue!) draft RFD on using LLMs at @oxide.computer, but I know that there is a ton more to be said on the topic; thoughts and experiences welcome!
rfd.shared.oxide.computer/rfd/0576
rfd.shared.oxide.computer/rfd/0576
576 - Using LLMs at Oxide / RFD / Oxide
rfd.shared.oxide.computer
December 7, 2025 at 9:08 AM
Good guide if you're needing to create policy for use of LLMs at your org (which you should)
When working with Claude Code I feel like I spend more time fettling instructions and md files than I do actually doing any work
December 5, 2025 at 5:28 PM
When working with Claude Code I feel like I spend more time fettling instructions and md files than I do actually doing any work
Reposted by Rob Bowley
A new academic paper from CMU analysing ~800 OSS repos shows AI is still making code worse.
AI speeds up code generation briefly, but code quality degrades quickly. Complexity rises faster than the amount of code added.
I wrote up the findings and my thoughts 👇
AI speeds up code generation briefly, but code quality degrades quickly. Complexity rises faster than the amount of code added.
I wrote up the findings and my thoughts 👇
AI Is still making code worse: A new CMU study confirms | Rob Bowley
blog.robbowley.net
December 4, 2025 at 12:19 PM
A new academic paper from CMU analysing ~800 OSS repos shows AI is still making code worse.
AI speeds up code generation briefly, but code quality degrades quickly. Complexity rises faster than the amount of code added.
I wrote up the findings and my thoughts 👇
AI speeds up code generation briefly, but code quality degrades quickly. Complexity rises faster than the amount of code added.
I wrote up the findings and my thoughts 👇
A new academic paper from CMU analysing ~800 OSS repos shows AI is still making code worse.
AI speeds up code generation briefly, but code quality degrades quickly. Complexity rises faster than the amount of code added.
I wrote up the findings and my thoughts 👇
AI speeds up code generation briefly, but code quality degrades quickly. Complexity rises faster than the amount of code added.
I wrote up the findings and my thoughts 👇
AI Is still making code worse: A new CMU study confirms | Rob Bowley
blog.robbowley.net
December 4, 2025 at 12:19 PM
A new academic paper from CMU analysing ~800 OSS repos shows AI is still making code worse.
AI speeds up code generation briefly, but code quality degrades quickly. Complexity rises faster than the amount of code added.
I wrote up the findings and my thoughts 👇
AI speeds up code generation briefly, but code quality degrades quickly. Complexity rises faster than the amount of code added.
I wrote up the findings and my thoughts 👇
OpenAI's "Code Red"
Traffic down 6% since Google launched new Gemini and Banana models
Traffic down 6% since Google launched new Gemini and Banana models
December 2, 2025 at 7:21 PM
OpenAI's "Code Red"
Traffic down 6% since Google launched new Gemini and Banana models
Traffic down 6% since Google launched new Gemini and Banana models
"The [Bank of England]'s financial stability report warned valuations are "particularly stretched" for companies focused on AI."
UK share values ‘most stretched’ since 2008, Bank warns
The bank warned US stock price valuations are their most stretched since the dotcom bubble burst.
www.bbc.co.uk
December 2, 2025 at 10:00 AM
"The [Bank of England]'s financial stability report warned valuations are "particularly stretched" for companies focused on AI."
Reposted by Rob Bowley
Headline: "AI can replace 11.7% of workforce"
Actual study: Anthropic paid MIT to use a "labor simulation tool" that said 11.7% of TASKS could be done by AI
Actual study: Anthropic paid MIT to use a "labor simulation tool" that said 11.7% of TASKS could be done by AI
MIT study finds AI can already replace 11.7% of U.S. workforce
Artificial intelligence can already replace 11.7% of the U.S. labor market, across finance, health care and professional services, according to MIT's study.
www.cnbc.com
November 29, 2025 at 9:57 PM
Headline: "AI can replace 11.7% of workforce"
Actual study: Anthropic paid MIT to use a "labor simulation tool" that said 11.7% of TASKS could be done by AI
Actual study: Anthropic paid MIT to use a "labor simulation tool" that said 11.7% of TASKS could be done by AI
"The question is, why haven’t we applied what we have repeatedly been forced to learn?"
Very long, but very good article.
Very long, but very good article.
Trillions Spent and Big Software Projects Are Still Failing
AI won’t solve IT’s management problems
spectrum-ieee-org.cdn.ampproject.org
November 29, 2025 at 7:12 PM
"The question is, why haven’t we applied what we have repeatedly been forced to learn?"
Very long, but very good article.
Very long, but very good article.
As a technologist I wholeheartedly endorse this message
User needs first
Service design second
Operating model third
Technology fourth
This order is still reversed way too often.
Service design second
Operating model third
Technology fourth
This order is still reversed way too often.
November 29, 2025 at 5:10 PM
As a technologist I wholeheartedly endorse this message