Jeffrey P. Bigham

model capability jaggedness is part of the unintuitiveness of current AI, but it's made even less intuitive by tirelessness… wigguming through the jaggedness toward something that looks like success. not quite a paperclip factory, but not so far off. metaphor for life? openai.…

openai.com

usually a fan of Zeynep's accounts, but AI will get better and better at customer service until it's at least as good as humans, … a big part of that will be systems engineering, including architectures that enable one AI interaction to share with the next www.nytimes.com/2026…

Opinion | Why A.I. Won’t Steal All Our Jobs - The New York Times nytimes.com

Apple announced a number of great new accessibility features today (better image description and q&a in voiceover, much more natural language voice control, and AI updates to accessibility reader) -- proud to have been part of delivering many of them! www.apple.com/newsroom/20…

Apple unveils new accessibility features, and updates with Apple Intelligence - Apple apple.com

it's not like Mercor is the first to do this, but interesting that AI companies hiring experts is (finally) getting attention :: www.nytimes.com/2026/07/10/b...

nytimes.com

anthropic has made some amazing products; they are also high on their own supply www.wired.com/story/anthro...

Anthropic Thinks Its Own Success Is Key to Making AI Safe wired.com

good to see this focus on asking clarifying from claude, lots of interesting questions about how to manage human agency while benefiting from automation -- choices is a direction we've been pursuing, e.g., in Morae and CowPilot -- www.cs.cmu.edu/~jbigham/pub... aclanthology.or…

cs.cmu.edu

good to see this focus on asking clarifying from claude, lots of interesting questions about how to manage human agency while benefiting from automation -- choices is a direction we've been pursuing, e.g., in Morae and CowPilot -- www.cs.cmu.edu/~jbigham/pub... aclanthology.or…

CowPilot: A Framework for Autonomous and Human-Agent Collaborative Web Navigation aclanthology.org

near where i grew up, somebody set a reminder to check back in 2046 --> "OpenAI agreed to a 20-year lease for the site, which will provide it eight gigawatts of computing capacity, or enough electricity to power about six million households in the United States." www.nytimes.c…

nytimes.com

what do we think, 3 minutes until someone makes a "claude watermark breaking" tool, "hey claude, break your watermark", thanks. techcrunch.com/2026/08/12/s...

Some Claude users are mad that Anthropic's new watermarks will catch them using it at their jobs, classes | TechCrunch techcrunch.com

what if every character was a captcha … www.creativebloq.com/design/fonts...

This free font creativebloq.com

2021->2022 felt like massive improvement, been pretty incremental since then, as people have improved models some, but also improved the harnesses and tools futurism.com/artificial-i...

Sam Altman Announces That the Singularity Has Arrived futurism.com

wild a $250Billion data center is being planned for tiny Piketon, OH -- this is near'ish my parents' place, ran a 5k there some summers, $250B could probably buy the whole place, wonder how much of that money they'll see, wonder how much it'll spike electricity prices -- qz.co…

qz.com

AI reviews are especially annoying because they are great at raising issues that appear to be legitimate to overworked associate editors, but which are in fact nonsense.

a lot of young folks are contacting me wanting to work on "human-AI alignment" and i think they just mean HCI, and the messaging battle has been lost.

in response to the flood of nonsense papers, too much of reviewing has become, "gotcha!", you used AI or you messed up a rule, desk reject. boom. science.

as confusing and chaotic as AI is right now, we'll be looking back at right now as the best days. the AI companies are competing for users, it's easy to access and mostly it's getting better. i can look up advice on products and it's mostly legit. when will the enshittification start?? :(

wouldn't it be neat to see where your AI computations happened, how many computers it involved, maybe even amount of power used … i imagine that information is harder than I might think to have access to real-time, but wouldn't it be cool?

there are far too many "AI" conferences being started.

I think about appropriate AI assistance in writing based on what human workflows are supported. from that way of thinking, it's surprising to me that NeurIPS used Pangram to reject papers that were "AI generated" b/c of the legit use cases it rejects -- consider bullet points to paragraph, e.g.,

a fascinating quality of current AI is its ability to take spoken ramblings and turn them into prose or action. speech has a lot of nice affordances, and yet it's been super hard to do much with it before because part of what makes it so easy means that it is imprecise and messy. AI smooths it out.

an underappreciated thing is that AI-written (or deeply rewritten) text is better than most people's actual writing. not everyone, and mostly because most people aren't that good at writing. and, my sense is this is worsening as people are being trained to like the AI writing more and more.