rolgalan
@rolgalan.bsky.social
I usually post about engineering processes, development productivity and anything related to software industry, with a particular interest in Android.
This is a wonderful parable explaining why some decisions cannot be made just staying in the middle of two options.
Particularly if these two options represent two fundamentally opposed world models whose average is simply the worst solution.
www.lesswrong.com/posts/LzQtrH...
Particularly if these two options represent two fundamentally opposed world models whose average is simply the worst solution.
www.lesswrong.com/posts/LzQtrH...
The Parable of the King and the Random Process — LessWrong
When advisors disagree wildly about when the rains will come, the king tries to average their predictions. His advisors explain why this is a terribl…
www.lesswrong.com
November 16, 2025 at 10:49 AM
This is a wonderful parable explaining why some decisions cannot be made just staying in the middle of two options.
Particularly if these two options represent two fundamentally opposed world models whose average is simply the worst solution.
www.lesswrong.com/posts/LzQtrH...
Particularly if these two options represent two fundamentally opposed world models whose average is simply the worst solution.
www.lesswrong.com/posts/LzQtrH...
«Language models hallucinate because standard training and evaluation procedures reward guessing over acknowledging uncertainty.»
«Most evaluations measure model performance in a way that encourages guessing rather than honesty about uncertainty.»
openai.com/index/why-la...
«Most evaluations measure model performance in a way that encourages guessing rather than honesty about uncertainty.»
openai.com/index/why-la...
Why language models hallucinate
OpenAI’s new research explains why language models hallucinate. The findings show how improved evaluations can enhance AI reliability, honesty, and safety.
openai.com
November 15, 2025 at 11:26 AM
«Language models hallucinate because standard training and evaluation procedures reward guessing over acknowledging uncertainty.»
«Most evaluations measure model performance in a way that encourages guessing rather than honesty about uncertainty.»
openai.com/index/why-la...
«Most evaluations measure model performance in a way that encourages guessing rather than honesty about uncertainty.»
openai.com/index/why-la...
«Like technical debt, management debt is incurred when you make an expedient, short-term management decision with an expensive, long-term consequence. If you incur the management debt without accounting for it, then you will eventually go management bankrupt.»
a16z.com/management-d...
a16z.com/management-d...
Management Debt | Andreessen Horowitz
When you base your life on credit and your loving days are done checks you signed with love and kisses later come back signed insufficient funds —Funkadelic, Can You Get to That Thanks to Ward Cunning...
a16z.com
November 15, 2025 at 11:08 AM
«Like technical debt, management debt is incurred when you make an expedient, short-term management decision with an expensive, long-term consequence. If you incur the management debt without accounting for it, then you will eventually go management bankrupt.»
a16z.com/management-d...
a16z.com/management-d...
«Geopolitical power once flowed through armies and treaties, but today it courses through silicon wafers, server farms and algorithmic systems. These invisible digital infrastructures and architectures shape every aspect of modern life.»
www.noemamag.com/reclaiming-e...
www.noemamag.com/reclaiming-e...
Reclaiming Europe’s Digital Sovereignty | NOEMA
Europe can accept permanent technological dependency, or it can build democratic digital systems rooted in climate commitments, labor protections and social diversity.
www.noemamag.com
November 9, 2025 at 5:49 PM
«Geopolitical power once flowed through armies and treaties, but today it courses through silicon wafers, server farms and algorithmic systems. These invisible digital infrastructures and architectures shape every aspect of modern life.»
www.noemamag.com/reclaiming-e...
www.noemamag.com/reclaiming-e...
«To move fast, you need to be autonomous and figure out who you need to influence directly for whatever purpose you have. You simply do not have the luxury of waiting around for your manager to use their authority and solve your problem.»
www.thecaringtechie.com/p/why-influe...
www.thecaringtechie.com/p/why-influe...
Why influence without authority is a vital skill for anyone
Whether you're an individual contributor or a leader, through proper influence you can get people to voluntarily comply, concede, or change. Who doesn't want that?
www.thecaringtechie.com
November 7, 2025 at 5:54 PM
«To move fast, you need to be autonomous and figure out who you need to influence directly for whatever purpose you have. You simply do not have the luxury of waiting around for your manager to use their authority and solve your problem.»
www.thecaringtechie.com/p/why-influe...
www.thecaringtechie.com/p/why-influe...
«The end result of all this "slowness" at the micro level is acceleration and efficiency at the macro level, and improved overall performance.»
www.lesswrong.com/posts/4FZfzq...
www.lesswrong.com/posts/4FZfzq...
"Slow is smooth, and smooth is fast" — LessWrong
I think it's worthwhile to signal boost bits of wisdom that turn out to be surprisingly useful. Apparently "Slow is smooth, and smooth is fast" is a…
www.lesswrong.com
November 2, 2025 at 11:33 AM
«The end result of all this "slowness" at the micro level is acceleration and efficiency at the macro level, and improved overall performance.»
www.lesswrong.com/posts/4FZfzq...
www.lesswrong.com/posts/4FZfzq...
«With Service Levels we try to shift the conversation from: "It should never fail (we don’t take any risk)" to "What is the definition of risk: SLI (risk assessment)" and "How much risk do we tolerate: SLO (risk management)”.»
blog.alexewerlof.com/p/why-bother...
blog.alexewerlof.com/p/why-bother...
Why bother with SLI and SLO?
Is there really any value in setting service level indicators and objectives?
blog.alexewerlof.com
November 2, 2025 at 11:27 AM
«With Service Levels we try to shift the conversation from: "It should never fail (we don’t take any risk)" to "What is the definition of risk: SLI (risk assessment)" and "How much risk do we tolerate: SLO (risk management)”.»
blog.alexewerlof.com/p/why-bother...
blog.alexewerlof.com/p/why-bother...
Observability on documentation! 🤯 Or some ideas to explore the value of technical docs.
When documentation is used? Does it correlates with a faster integration with your service? Does it increase the usage of the tool you provide?
If only was easy to measure this...
passo.uno/docs-observa...
When documentation is used? Does it correlates with a faster integration with your service? Does it increase the usage of the tool you provide?
If only was easy to measure this...
passo.uno/docs-observa...
Docs observability, or measuring docs inside a product-docs system
As technical writers we want to know if the docs we’re writing are accomplishing their goals. In other words, we want to know how good are docs relative to the business goals they’re aiming to support...
passo.uno
November 1, 2025 at 11:04 AM
Observability on documentation! 🤯 Or some ideas to explore the value of technical docs.
When documentation is used? Does it correlates with a faster integration with your service? Does it increase the usage of the tool you provide?
If only was easy to measure this...
passo.uno/docs-observa...
When documentation is used? Does it correlates with a faster integration with your service? Does it increase the usage of the tool you provide?
If only was easy to measure this...
passo.uno/docs-observa...
Reposted by rolgalan
This is a great research assessing the Quality and Security of AI-Generated Code published just a few months ago.
It proves that quality of code generated by LLMs is quite poor, and what is even more dangerous: it might look like it’s correct, but still contain bugs.
arxiv.org/abs/2508.14727
It proves that quality of code generated by LLMs is quite poor, and what is even more dangerous: it might look like it’s correct, but still contain bugs.
arxiv.org/abs/2508.14727
Assessing the Quality and Security of AI-Generated Code: A Quantitative Analysis
This study presents a quantitative evaluation of the code quality and security of five prominent Large Language Models (LLMs): Claude Sonnet 4, Claude 3.7 Sonnet, GPT-4o, Llama 3.2 90B, and OpenCoder ...
arxiv.org
October 26, 2025 at 1:12 PM
This is a great research assessing the Quality and Security of AI-Generated Code published just a few months ago.
It proves that quality of code generated by LLMs is quite poor, and what is even more dangerous: it might look like it’s correct, but still contain bugs.
arxiv.org/abs/2508.14727
It proves that quality of code generated by LLMs is quite poor, and what is even more dangerous: it might look like it’s correct, but still contain bugs.
arxiv.org/abs/2508.14727
Working in a platform central team that supports other engineers means answering questions frequently.
There are many resources about 'How to ask good questions'. However is not so frequent to read tips about what is a good answer. Find some good examples here!
jensrantil.github.io/posts/a-good...
There are many resources about 'How to ask good questions'. However is not so frequent to read tips about what is a good answer. Find some good examples here!
jensrantil.github.io/posts/a-good...
What's in a good answer?
Answering questions the right way can dodge many questions in the future.
jensrantil.github.io
October 26, 2025 at 12:17 PM
Working in a platform central team that supports other engineers means answering questions frequently.
There are many resources about 'How to ask good questions'. However is not so frequent to read tips about what is a good answer. Find some good examples here!
jensrantil.github.io/posts/a-good...
There are many resources about 'How to ask good questions'. However is not so frequent to read tips about what is a good answer. Find some good examples here!
jensrantil.github.io/posts/a-good...
Reposted by rolgalan
Whether you’re an AI skeptic, an AI evangelist, or just someone who wants to know what’s happening on the cutting edge of tech, you should be reading @garymarcus.bsky.social.
Vibe coding? It’s dying. I am shocked. SHOCKED, I tell you.
garymarcus.substack.com/p/is-vibe-co...
Vibe coding? It’s dying. I am shocked. SHOCKED, I tell you.
garymarcus.substack.com/p/is-vibe-co...
Is vibe coding dying?
Amateurs might not be replacing teams of coders, after all
garymarcus.substack.com
October 22, 2025 at 4:38 PM
Whether you’re an AI skeptic, an AI evangelist, or just someone who wants to know what’s happening on the cutting edge of tech, you should be reading @garymarcus.bsky.social.
Vibe coding? It’s dying. I am shocked. SHOCKED, I tell you.
garymarcus.substack.com/p/is-vibe-co...
Vibe coding? It’s dying. I am shocked. SHOCKED, I tell you.
garymarcus.substack.com/p/is-vibe-co...
«Ad platforms are selling impressions to bots. Businesses are buying fake traffic to inflate their metrics. Analytics companies are dutifully reporting on this bot activity. And the entire industry seems to be nodding along, complicit in a collective charade.»
joindatacops.com/resources/ho...
joindatacops.com/resources/ho...
How 73% of Your E-commerce Visitors Could Be Fake
A conversion rate of less than 0.1%. That was the moment I realized something was fundamentally broken with the way we measure success on the internet.
joindatacops.com
October 18, 2025 at 8:03 AM
«Ad platforms are selling impressions to bots. Businesses are buying fake traffic to inflate their metrics. Analytics companies are dutifully reporting on this bot activity. And the entire industry seems to be nodding along, complicit in a collective charade.»
joindatacops.com/resources/ho...
joindatacops.com/resources/ho...
Best tips for library maintainers: use the latest toolchain, but make sure you compile and ship a library backwards compatible to older versions of java and kotlin.
A guide for setting up a #gradle for a library successfully.
blog.alllex.me/posts/2023-1...
A guide for setting up a #gradle for a library successfully.
blog.alllex.me/posts/2023-1...
Conservative libraries with liberal tooling
How to reach more users for Kotlin and Java libraries, yet keep the modern tooling
blog.alllex.me
October 12, 2025 at 7:03 PM
Best tips for library maintainers: use the latest toolchain, but make sure you compile and ship a library backwards compatible to older versions of java and kotlin.
A guide for setting up a #gradle for a library successfully.
blog.alllex.me/posts/2023-1...
A guide for setting up a #gradle for a library successfully.
blog.alllex.me/posts/2023-1...
These are astonishing figures for something that is not able to give 2 answers correct in a row.
«The industry probably needs a revenue range that is closer to the $320 billion to $480 billion range, just to break even on the capex to be spent this year 2025.»
futurism.com/future-socie...
«The industry probably needs a revenue range that is closer to the $320 billion to $480 billion range, just to break even on the capex to be spent this year 2025.»
futurism.com/future-socie...
AI Data Centers Are an Even Bigger Disaster Than Previously Thought
An investment manager realized he made a crucial mistake — and that his grim prediction about AI investments may not have been cynical enough.
futurism.com
October 12, 2025 at 6:23 PM
These are astonishing figures for something that is not able to give 2 answers correct in a row.
«The industry probably needs a revenue range that is closer to the $320 billion to $480 billion range, just to break even on the capex to be spent this year 2025.»
futurism.com/future-socie...
«The industry probably needs a revenue range that is closer to the $320 billion to $480 billion range, just to break even on the capex to be spent this year 2025.»
futurism.com/future-socie...
This is a quite complete article about what it means to be a staff+ engineer.
«Staff Engineers operate over longer time horizons. One could say that the Staff’s role is more strategic whereas the other roles are more tactical.»
blog.alexewerlof.com/p/senior-eng...
«Staff Engineers operate over longer time horizons. One could say that the Staff’s role is more strategic whereas the other roles are more tactical.»
blog.alexewerlof.com/p/senior-eng...
Senior to Staff Engineer
What are the similarities and differences? What are the criteria for promotion? And some tools and techniques that come in handy.
blog.alexewerlof.com
October 12, 2025 at 5:25 PM
This is a quite complete article about what it means to be a staff+ engineer.
«Staff Engineers operate over longer time horizons. One could say that the Staff’s role is more strategic whereas the other roles are more tactical.»
blog.alexewerlof.com/p/senior-eng...
«Staff Engineers operate over longer time horizons. One could say that the Staff’s role is more strategic whereas the other roles are more tactical.»
blog.alexewerlof.com/p/senior-eng...
This is exactly the work that platform teams need to do to maximize the developer productivity with faster builds: Producing telemetry, cleaning outliers, categorizing type of builds, comparing different scenarios and doing a solid analysis to reach the final conclusion.
incident.io/blog/festive...
incident.io/blog/festive...
Tracking developer build times to decide if the M3 MacBook is worth upgrading | Blog | incident.io
When our CTO said "I'll upgrade your MacBook if you can prove it's worthwhile", we embarked on a journey including (re)building a Go hot-reloader, instrumenting developer builds, analyzing compiler pe...
incident.io
October 4, 2025 at 6:48 AM
This is exactly the work that platform teams need to do to maximize the developer productivity with faster builds: Producing telemetry, cleaning outliers, categorizing type of builds, comparing different scenarios and doing a solid analysis to reach the final conclusion.
incident.io/blog/festive...
incident.io/blog/festive...
Quite in-deep analysis on LLM performance on chess, particularly trying to figure out why so much divergence between each model.
dynomight.net/more-chess/
dynomight.net/more-chess/
OK, I can partly explain the LLM chess weirdness now
(“make LLMs play better with one weird trick”)
dynomight.net
September 30, 2025 at 6:20 PM
Quite in-deep analysis on LLM performance on chess, particularly trying to figure out why so much divergence between each model.
dynomight.net/more-chess/
dynomight.net/more-chess/
Really insightful analysis of the impact of AI assisted code changes during the last years across several projects.
«This graph shows that for every 25% increase in the adoption of AI, their model projects a 7.2% decrease in “delivery stability»
www.gitclear.com/ai_assistant...
«This graph shows that for every 25% increase in the adoption of AI, their model projects a 7.2% decrease in “delivery stability»
www.gitclear.com/ai_assistant...
AI Copilot Code Quality: 2025 Data Suggests 4x Growth in Code Clones - GitClear
www.gitclear.com
September 28, 2025 at 2:40 PM
Really insightful analysis of the impact of AI assisted code changes during the last years across several projects.
«This graph shows that for every 25% increase in the adoption of AI, their model projects a 7.2% decrease in “delivery stability»
www.gitclear.com/ai_assistant...
«This graph shows that for every 25% increase in the adoption of AI, their model projects a 7.2% decrease in “delivery stability»
www.gitclear.com/ai_assistant...
Muy relevante este artículo con un mensaje muy claro: Hemos confundido la capacidad de generar código con la capacidad de entregar valor.
Esto ya sucedía antes de la IA generando código, pero utilizarla sin disciplina tiende a agravar el problema.
nikeyes.github.io/el-codigo-ge...
Esto ya sucedía antes de la IA generando código, pero utilizarla sin disciplina tiende a agravar el problema.
nikeyes.github.io/el-codigo-ge...
September 24, 2025 at 3:30 PM
Muy relevante este artículo con un mensaje muy claro: Hemos confundido la capacidad de generar código con la capacidad de entregar valor.
Esto ya sucedía antes de la IA generando código, pero utilizarla sin disciplina tiende a agravar el problema.
nikeyes.github.io/el-codigo-ge...
Esto ya sucedía antes de la IA generando código, pero utilizarla sin disciplina tiende a agravar el problema.
nikeyes.github.io/el-codigo-ge...
«Quality does not depend only on tools or individual talent. It is a direct consequence of the environment (system) we build. Quality does not arise spontaneously: it needs space, alignment, and a culture that values it.»
www.eferro.net/2025/06/over...
www.eferro.net/2025/06/over...
Lean Software Development: Overcoming resistance and creating conditions for quality
Fifth article on quality in Lean Software Development. In previous posts, we talked about how to build with quality through mistakes, techni...
www.eferro.net
September 21, 2025 at 9:47 AM
«Quality does not depend only on tools or individual talent. It is a direct consequence of the environment (system) we build. Quality does not arise spontaneously: it needs space, alignment, and a culture that values it.»
www.eferro.net/2025/06/over...
www.eferro.net/2025/06/over...
«It's not about “beautifying the code” or following arbitrary rules, but about building a solid technical foundation that reduces everyday friction and enables rapid, confident progress.»
www.eferro.net/2025/05/qual...
www.eferro.net/2025/05/qual...
Lean Software Development: Quality as the Foundation of Sustainable Development
Third part of the series on quality in Lean Software Development. After exploring how to detect errors early and learn from them, in this en...
www.eferro.net
September 21, 2025 at 8:57 AM
«It's not about “beautifying the code” or following arbitrary rules, but about building a solid technical foundation that reduces everyday friction and enables rapid, confident progress.»
www.eferro.net/2025/05/qual...
www.eferro.net/2025/05/qual...
Quite complete article about the reasons for non-determinism in LLMs, why is important and how it could be solved.
thinkingmachines.ai/blog/defeati...
thinkingmachines.ai/blog/defeati...
Defeating Nondeterminism in LLM Inference
Reproducibility is a bedrock of scientific progress. However, it’s remarkably difficult to get reproducible results out of large language models.
For example, you might observe that asking ChatGPT the...
thinkingmachines.ai
September 20, 2025 at 8:45 AM
Quite complete article about the reasons for non-determinism in LLMs, why is important and how it could be solved.
thinkingmachines.ai/blog/defeati...
thinkingmachines.ai/blog/defeati...
This is insane, a really simple task as copying URLs does not have 100% of success rate for most #LLMs. This is a really nice use case as a single character wrong might break the whole url.
Just consider the outcome of trying autonomous agents with this confidence level.
github.com/Mankind-Tech...
Just consider the outcome of trying autonomous agents with this confidence level.
github.com/Mankind-Tech...
September 17, 2025 at 6:54 PM
This is insane, a really simple task as copying URLs does not have 100% of success rate for most #LLMs. This is a really nice use case as a single character wrong might break the whole url.
Just consider the outcome of trying autonomous agents with this confidence level.
github.com/Mankind-Tech...
Just consider the outcome of trying autonomous agents with this confidence level.
github.com/Mankind-Tech...
Grandísima noticia hacia la transparencia digital en España. ¡Bravo! 🙌
civio.es/novedades/20...
civio.es/novedades/20...
Civio abre camino en la transparencia algorítmica: el Supremo condena al Gobierno a entregar el código fuente de BOSCO
La sentencia, que crea jurisprudencia, considera que el uso de aplicaciones en la toma de decisiones públicas debe ir acompañada de transparencia que permita su fiscalización. Rechaza que la seguridad...
civio.es
September 17, 2025 at 6:43 PM
Grandísima noticia hacia la transparencia digital en España. ¡Bravo! 🙌
civio.es/novedades/20...
civio.es/novedades/20...