zacharyozer

> Earlier this year, Atallah described OpenRouter as the AI equivalent of Stripe.

Not sure I understand how this is strategically aligned for Stripe but certainly an interesting comparison.

> According to Dangel, it costs $2,600 to run Anthropic’s LLM Opus 4.6 through RULER 128, a test developed by Nvidia to assess a model’s ability to retrieve information from large data sets. And SubQ? “It cost us eight dollars,” he says.

> SubQ does seem to be able to handle a lot of text at once. The model has a context window (roughly akin to a working memory) up to 12 million tokens long. Most top models today have context windows one million tokens long. In a demo that Whedon ran for me, he asked SubQ to perform a task that required it to reason about information contained in 400 documents. It responded in seconds. When he gave Perplexity—a popular LLM-powered search engine—the same task, it failed to load all 400 documents.

Congrats on the release, and love the write up!

I’m increasingly thankful Zulip exists as an open source alternative to other collaborative platforms. In the age of agents, it feels like a huge advantage that folks can take a robust platform as a base and customize it to their needs (and hopefully contribute back)!

Also, thanks for sharing about your experiences with agent-based contributions, both internally and externally. Super fascinating.

Be well!

Interesting baseline of how much AI can help with the profitability of a business:

> By the end of last year, Medvi had reached $401 million in annual sales and amassed 250,000 customers. It produced 16.2 percent in net profit, or $65 million, with spending going to the fees for telehealth platforms, marketing and then software. Hims, by contrast, had a net profit of 5.5 percent last year.

> Manufacturing job share has been dwindling in nearly all middle- and high-income countries. By one analysis, China lost more than 30 million of those jobs from 2011 to 2020, more than twice as many as the number of jobs that exist in the entire U.S. manufacturing sector. Yet total employment continued to grow, meaning the relative share of jobs in manufacturing fell even more sharply than the numbers alone would indicate.

> Manufacturing output, meanwhile, has risen, because workers now produce far more per hour using better and more sophisticated equipment. Today a given number of autoworkers can make, according to my calculations, three times as many cars in a year as they could 50 years ago.

> The problem is that consumers do not want three times as many cars. Even as people get richer, they increase their spending on manufactured goods only modestly, preferring instead to spend more on services like travel, health care and dining out. There are only so many cars a family can own, but that’s not the case for expensive vacations or fancy meals. As a result we have fewer people working in auto factories and more people working in luxury resorts and the like.

Congrats Steve! I haven't touched video since I was at JW Player a million years ago, but I always inspired by the simplicity of video.js (especially the theming).

Hope this new iteration is exceptionally successful.

Co-Author of TaxCalcBench here.

The advice is correct: don’t use AI to file your taxes!

But what the article misses is that the uplift we’ve seen across models is incredible. Don’t be surprised if next year the headline reads “A.I. newest trick: Filing your taxes”.