this is why it's so critical (IMO) to find ways to tune the models to produce more out-of-distribution outputs. it's incredibly easy to generate "in-distribution" text and the major labs are optimizing for this because of "safety", but the only way to generate truly creative outputs is to step in and out of the fringe.
Interesting! My first thought looking at the scribble chart was "isn't this Monte Carlo simulation?" but reading further it seems more aligned with the "third way" that William Briggs describes in his book Uncertainty[1]. He argues we should focus on direct probability statements about observables over getting lost in parameter estimation or hypothesis testing.
Sole prop registrations (for individuals/non-corporations) are not _too_ onerous anymore and generally have a faster time to approval, provided you submit the right docs. Totally agree it's a money grab though. Twilio et al. have so much power to keep bad actors at bay but keep pushing responsibility to the consumers of their API. I feel like setting up push notifications is easier though, which is saying something.
Everything old is new again: I came across a demo for Telescript [1] the other day that would not look out of place in a pitch deck today, save the references to AT&T. https://www.youtube.com/watch?v=wtrs3jtY96k
As a potential solution, do you think formal/semi-formal software development education (undergrad programs, colleges/polytechnics, dev bootcamps, etc) should lean super heavily into AI? To the extent that it's not just "use ChatGPT to help you complete this assignment" but rather "complete this assignment using *only* ChatGPT: you're not allowed to write any of the code by yourself".
What I've noticed from my extensive use over the past couple weeks has been Claude Code really sucks at thinking things through enough to understand the second and third order consequences of the code that it's writing. That said, it's easy enough to work around its deficiencies by using a model with extended thinking (Grok, GPT4.5, Sonnet 3.7 in thinking mode) to write prompts for it and use Claude Code as basically a dumb code-spewing minion. My workflow has been: give Grok enough context on the problem with specific code examples, ask it to develop an implementation plan that a junior developer can follow, and paste the result into Claude Code, asking it to diligently follow the implementation plan and nothing else.
Libraries like iText would be SO good with LLM/vision model integration and vice-versa. Huge opportunity to use these tools to generate more training data based from siloed PDFs.