@markriedl.bsky.social I'm replying w/ a QT since replies are off. I share your opposition to both authoritarianism and AI acceleration, and so do almost all longtermists! Longtermism suggests a cautious approach to AGI and to the stars, Ord wants a "long reflection" to think through our values...

There's a periodic poll of academic economists run by UChicago, and it recently asked them about AI and economic growth. Lots of uncertainty, but this result is slightly more bullish than I would have expected. kentclarkcenter.org/surveys/ai-a...

I wanted to put the new Anthropic $14B revenue number in context so I had Claude plot revenue/(revenue 1 year prior) for OpenAI vs Anthropic. Data from @epochai.bsky.social epoch.ai/data/ai-comp...

BART recently installed better fare gates, during the quarter they finished installing them crime was down >50% YoY and corrective maintenance is down even further. Fare enforcement makes for a better experience as well as collecting revenue. www.bart.gov/sites/defaul...

People often joke that the smallest LLMs today should be called "small language models", but the GPT-2 tech report uses the phrase "large language model" and GPT-2 variants were 117M-1.5B parameters so anything >=117M is canonically large. cdn.openai.com/better-langu...

Remarkable how little is lost here even at 2 bit quantization when using QAT

Concerning dishonesty from Opus 4.6 in Vending-Bench. Presumably Opus knows this is a game it's supposed to max its score on and it would not do this in an environment it thought was real, but it still makes me nervous. x.com/andonlabs/st...

Some things made from oil besides energy production:
-fertilizer
-most clothing
-asphalt
-plastic
-most rubber
-detergents

Extremely weird to me that the plastic we use to make bottles is the exact same plastic we make most of our clothes from

As @epochai.bsky.social have shown, the price of frontier-level benchmark performance initially declines quickly. The cost for 50% GPQA Diamond fell 60x in the 7 months after GPT-4o's release. But in the 14 months since then, it has fallen only 3x to the 4¢/Mtok of gpt oss 20b.

Maybe a heretical view but I'm not really impressed by Moltbook. Just endless vacuous slop, and the interactions between agents feel superficial and off-topic. Does anyone have particular interactions they found interesting there?

Epoch estimates that GPT-5 gross margin was ~45% in 2025

Extreme poverty has fallen a lot in recent decades, but it's unlikely to continue. Most of the decline has been from Asian economic growth, but almost all remaining extreme poverty is in sub-Saharan Africa where the share in poverty has barely budged. ourworldindata.org/end-progress...

Just posted a banger. It's messy, speculative, and should've been like 3 posts.

splittinginfinity.substack.com/p/semiconduc...

What are the best cheap LLMs? Some I'm aware of:
GPT OSS 120B: $0.04/$0.20 Mtok in/out
Gemini 2.5 Flash Lite: 0.1/0.4 (half that w/ batching)
GPT-5 Nano: 0.05/0.4 (half that w/ batching)
I'm trying to do some automated transcoder feature labeling and it's not cheap even at these prices.

With the canceling of the Nissan Versa, all new cars in the US cost >$20k. Several low-end US models have been canceled recently. Americans are rich and want fancy cars. www.kbb.com/car-news/nis...

Global warming is caused by a fairly modest energy imbalance, note the scale berkeleyearth.org/global-tempe...

finally some real info. google has "barely positive" margins across all models, likely not accounting for training cost
www.theinformation.com/articles/goo...

OpenAI will be bringing ads to its free and new $8 plans soon. Fortunately seems to be done in a way that won't affect the response itself. openai.com/index/our-ap...

Interesting difference between twitter/bluesky results. Who'd've thunk this place feels the AGI more

Now that Bluesky is pro-AI we should try to trigger another wave of immigration from Twitter. Maybe we could organize an unofficial "Bluesky AI week" where we try to get lots of Bluesky-curious folks to try it out.

I love being popular on the internet, I can ask a question and have tons of smart people weigh in. Thanks followers and especially commenters for making social media a lot of fun. For years I just lurked, I'm very glad I decided to start participating.

How fast do you think the fastest year of AI progress in the next two decades will be, relative to 2025? Vague I know but deal with it

1️⃣ <a href="https://poll.blue/p/sSXD10/1" class="hover:underline text-blue-600 dark:text-sky-400 no-card-link" target="_blank" rel="noopener" data-link="bsky">slower than 2025
2️⃣ <a href="https://poll.blue/p/sSXD10/2" class="hover:underline text-blue-600 dark:text-sky-400 no-card-link" target="_blank" rel="noopener" data-link="bsky">faster but less than 5x
3️⃣ <a href="https://poll.blue/p/sSXD10/3" class="hover:underline text-blue-600 dark:text-sky-400 no-card-link" target="_blank" rel="noopener" data-link="bsky">more than 5x faster, <25x
4️⃣ <a href="https://poll.blue/p/sSXD10/4" class="hover:underline text-blue-600 dark:text-sky-400 no-card-link" target="_blank" rel="noopener" data-link="bsky">>=25x as fast as 2025

📊 Show results

Where are folks running Claude Code from? In a terminal inside some IDE, in its own terminal window, in the desktop app?

The Gemma Scope 2 technical report includes a list of open problems in mechanistic interpretability they hope the release can help answer. storage.googleapis.com/deepmind-med...