Daniel Bachler

Daniel Bachler

@danielbachler.de

Head of Engineering at @ourworldindata.org; likes nuanced discourse, weird creatures and programming languages with ADTs

1.3K followers887 following92 posts

Daniel Bachler reposted

The biohackers subreddit banned all posts about peptides and HRT because peptide companies were spamming + manipulating the sub in order to manipulate Google AI and ChatGPT answers. Very insidious industry that is drastically changing the internet www.404media.co/companies-ar...

404media.coCompanies Are Using Reddit to Manipulate ChatGPT and Google AI SearchPeptide companies have been doing AI-engine optimization by spamming the biohackers subreddit to manipulate ChatGPT and Google.

Daniel Bachler reposted

View full thread
View full thread

In her piece, Hannah varies the three key inputs to the simulation one by one: Total Fertility Rate: children per woman, on average Life Expectancy: how long the average person lives Net Migration: immigrants minus emigrants

The tool then plots the population curve and age pyramid through 2100. It's deliberately simple so it can run in real time in your browser — and even so, given the same inputs, it matches UN projections closely for most countries.

Things I find surprising about AI in Sept 2025: - lots of progress on coding agents, little on game-playing - really great writing is still beyond LLMs, no real progress since 4o - AI generated videos are really convincing now, but fake AI content is not flooding my timeline yet

Excellent thread on AI by my colleague Esteban - recommended!

Esteban Ortiz-Ospina@eortizospina.bsky.social

1/ I recently listened to David Autor on @reidhoffman.bsky.social's podcast and really enjoyed it. It got me thinking about how we've adapted to past technological shifts and why we need to think differently about adaptation with AI. A few thoughts in this thread🧵

So far we’ve been lucky that LLM economics results in almost perfectly egalitarian access to capabilities. It could’ve gone differently - where the richest enjoy far better systems or one lab holds a monopoly. Here’s hoping this stays true.

Really nice to see these features in prod! Kudos to my colleague @sophiamersmann.bsky.social for building all of this! 🔥

Our World in Data@ourworldindata.org

💡 Very excited to announce new features: better interactive maps! 🗺️ – Enhanced country/region selection with several sorting options and a bar chart preview – Selecting a country highlights it and shows its value on the map – Zooming to your selection brings up a 3D globe view 🌍

Daniel Bachler reposted

“…today’s antivirals are so effective that if taken during pregnancy, the risk of passing HIV to the infant is minimal. Without these drugs, around a quarter of babies born to HIV-positive mothers would also contract the virus – and half of them would die before the age of two.”

Saloni@scientificdiscovery.dev

I signed the 10% Giving Pledge this year, and this week, I decided to donate 19% of my annual income to global health charities. It's not going to be anywhere near "enough", but I'm happy to do it --

Leopold Aschenbrenner and Daniel Kokotajlo have both argued that China must not reach AGI first. Have either of them given any update on how they think about US labs reaching AGI while Trump is in the White House?

Something pretty strange about where we are with LLMs is how bad the spam in my email inbox still is. I would have guessed we would routinely see super convincing, personalized spam by now.

Daniel Bachler reposted

Tell me your team is entirely comprised of over-confident twenty-somethings without telling me your team is entirely comprised of over-confident twenty-somethings

Katie Drummond@katie-drummond.bsky.social

NEW: DOGE is planning to migrate the Social Security Administration codebase to a new programming language. This would normally take years. DOGE wants it done in months. That could put benefits for millions of Americans at risk. Scoop from @makenakelly.bsky.social www.wired.com/story/doge-r...

I find many non-fiction books are more verbose than I would like. I'd often prefer a distilled version with a page count of roughly a quarter. I'd be willing to pay more for the distilled version than the verbose one. Could this be a new revenue stream for writers in the age of AI?

Daniel Bachler reposted

Our latest issue of Works in Progress dropped today! - The steam networks of NYC - Prehistoric violence - Urbanism with Chinese characteristics - Extending the fertility window - The Hanseatic League's rise and fall - The pineapple: the king of fruits - The land value tax worksinprogress.co

Daniel Bachler reposted

Most OECD countries fail to allocate even 0.15% of their national income to aid the world’s poorest countries, despite a commitment to do so. That's just $1 out of every $700. Only Luxembourg, Sweden, and Norway meet this minimal UN target. New cuts will deepen this shortfall.

Daniel Bachler reposted

South Korea’s under-15 population has halved in two generations. I knew the population was aging quickly, and it’s no surprise that means fewer children—but the absolute numbers are striking. It's easy to look at fertility rates and forget about absolute numbers. That's a lot fewer kids around.

ourworldindata.orgThe number of children in South Korea has fallen by 60% since its peakSouth Korea is undergoing one of the world’s most rapid demographic transitions. Fertility rates — the number of children a woman has over her lifetime — have fallen rapidly over the last 50 years, an...

I think this is right. There is another interesting dynamic that for AI labs it would be very useful if academic research were freely available and so you have a powerful new force against established, paywalled scientific publishing.

Esteban Ortiz-Ospina@eortizospina.bsky.social

I think AI tools like DeepResearch will break the market of peer-reviewed, paywalled journals sooner than most people realize. Evaluation is harder than production of papers, and that gap is growing quickly

OpenAI plus users now have access to the "Deep Research" feature. I just tried it for the first time and asked it for a report on a niche programming topic (comparing computation expressions in F# vs Algebraic Effects in OCaml) and the result was better than any blog post I know of on the topic.

This is quite surprising!

Ethan Mollick@emollick.bsky.social

This is a crazy paper. Fine-tuning a big GPT-4o on a small amount of insecure code or even "bad numbers" (like 666) makes them misaligned in almost everything else. They are more likely to start offering misinformation, spouting anti-human values, and talk about admiring dictators. Why is unclear.

Daniel Bachler reposted

AMFAR has published a report on how the foreign aid freeze affects PEPFAR, which actually still hasn't resumed. They estimate that hundreds of thousands of people lose access to HIV antivirals and 1500 babies are expected to contract HIV — *each day* 🧵

Daniel Bachler reposted

NEW | Solar generates more EU electricity than coal for the first time in 2024 🇪🇺☀️ Renewables made up nearly HALF of the electricity mix last year, while fossil power fell to a historic low ⚡ ember-energy.org/lat...

Daniel Bachler reposted

Simon is one of the best voices on AI on the internet and his predictions in this post seem very reasonable. I particularly share his scepticism that general purpose AI Agents will work well soon. Interesting times ahead in any case!

Simon Willison@simonwillison.net

Risky post... here are my 1, 3 and 6 year predictions for LLMs/AI, expanded from my appearance on the Oxide and Friends podcast (My confidence in my ability to predict the future is extremely low) simonwillison.net/2025/Jan/10/...

Nathan wrote a great post about what we know so far about o3, the recently announced new LLM by OpenAI

Nathan Lambert@natolambert.bsky.social

OpenAI's o3: The grand finale of AI in 2024 A step change as influential as the release of GPT-4. Reasoning language models are the current and next big thing. I explain: * The ARC prize * o3 model size / cost * Dispelling training myths * Extreme benchmark progress

A lot of use cases could be unlocked if prompt injections and jailbreaking could be prevented entirely, but even making them very very hard to pull off would already be super valuable.

Amazing analysis! A fast rollout of malaria vaccines would be incredible - let’s do this!

Justin Sandefur@justsand.bsky.social

New @cgdev.org policy paper: *MALARIA VACCINES* are a huge scientific breakthrough. But the world is dragging its feet on rolling them out. We examine why, and how to go faster. 🧵

If you are using obsidian then do yourself a favour and add the web clipper extension to your browser. It makes getting content from the web into obsidian much nicer and has replaced bookmarks for many use-cases for me. Inline LLM support is a great new addition but it’s really good without as well.

kepano@stephango.com

The latest version of Obsidian Web Clipper adds Interpreter which lets you extract page content and manipulate it using natural language. This is super useful to create consistent data structures for research within your vault. You can run it locally using Ollama. help.obsidian.md/web-clipper/...

Daniel Bachler reposted

The world is awful. The world is much better. The world can be much better. All three statements are true at the same time. Understanding this is key to solving big global problems. We believe data & research can help us understand both the problems we face & the progress that’s possible. 🧵

Daniel Bachler reposted

Good opportunity to point out a nice OWID feature: interactive charts have dynamic social media preview cards.

If you adjust the chart parameters (add countries, change timeframe, etc.), the URL updates automatically. Share the link, and the social card shows exactly the view you customized ✨

Esteban Ortiz-Ospina@eortizospina.bsky.social

“In 1800, 1 in 3 people on earth were Chinese. Today, it’s less than 1 in 5” From Tom Whitwell's "52 things I learned in 2024": medium.com/@tomwhitwell... h/t @maxroser.bsky.social