A big corporate misconception of Generative AI: Pre-LLM AI was training models on your data to find hidden insight. Firms are trying to do that with LLMs but they don't really work that way, they aren't analysis tools & your data matters less, they are pre-trained on the internet
@emollick
-
AI Companies Neglect Language and Reasoning Benchmarks for Software Optimization
By
–
There are so few benchmarks the AI companies compete on outside of software and general knowledge benchmarks. They also fine tune obsessively to optimize software. Language, turn-taking, logical reasoning, lack of hallucinations, and other critical issues get less clear focus.
-
AI Development Bias: Software Engineers Overlook Marketing, Education, Law
By
–
The Law of the Hammer is an issue in AI. The people who build AI are software engineers, and they obsess over the degree to which LLMs can (or cannot) automate software. Meanwhile, entire swathes of larger industries exposed to AI (marketing, education, law) get much less focus.
-
Social Media Election Influence: Facebook Voting Ads Impact
By
–
How can a social media company swing elections? We already know the answer as @jengolbeck pointed out: a 2012 study found Facebook banner ads reminding people to vote generated "340,000 additional votes" Companies don't need to show the ads to everyone… https://
nature.com/articles/natur
e11421
… -
Video Trend Critique: Social Media Content Strategy Concerns
By
–
Yes, yes that is exactly what I want on Twitter is to watch a bunch of videos like this one which takes 14 seconds to show the vital message "Video is the future. X Indispensable"
— Ethan Mollick (@emollick) 13 mars 2024
Looking forward to turning every paper I post into a 4 hour video of swirling words. Awesome. Great https://t.co/rhjtj4OFzyYes, yes that is exactly what I want on Twitter is to watch a bunch of videos like this one which takes 14 seconds to show the vital message "Video is the future. X Indispensable" Looking forward to turning every paper I post into a 4 hour video of swirling words. Awesome. Great
-
LLM Capabilities Growing Several Times Faster Than Moore’s Law
By
–
Here's a good estimate of how fast the capabilities of LLMs have been growing: several times as fast as Moore's Law! The compute needed to achieve the same outcome halving every 5 to 14 months, with no sign of slowing. Most gains are from bigger scale. https://
arxiv.org/abs/2403.05812 -
GPT-4 Demonstrates Naval Engineering Design Capabilities Through Simulation
By
–
GPT-4 has surprisingly good suggestions on how to design a 1890s-era armored cruiser, down to concerns about how the casement guns might limit operations in heavy seas. (I am giving it screenshots from a rather intensely detailed spreadsheet-like naval simulation game. Neat!)
-
LLMs Superior to Dictionary Methods for Psychological Text Analysis
By
–
LLMs are much better than the standard dictionary approaches to figuring out the psychological context of text, including uncommon African languages, and likely cheaper & easier-to-use as well It is approaching human levels, and is improving as models do. https://
osf.io/preprints/psya
rxiv/sekf5
… -
Government AI governance panel with defense and cyber leaders
By
–
I was lucky to be part of a panel led by former Defense Secretary Leon Panetta w/ Adm. Mike Roger (former head of Cyber Command and the NSA), Prof. @jengolbeck & @aneeshchopra (former US CTO). A lesson: government taking upsides & risks of AI seriously https://
youtu.be/MJ_8CHkPttY?si
=LMx_yXmpTT7xXBi8
… -
Devoted Users May Not Represent Wider Market Needs
By
–
Listening to your most devoted users can lead you wrong because they have their own specific needs, which may not represent the needs of the wider market. In fact, they don't want you to solve for the mass market – the classic "Chasm" for new products. I just saw this on Steam.