weekend ai reads for 2024-05-17

May 17, 2024

šŸ“° ABOVE THE FOLD: GOOGLE & OPENAI ANNOUNCEMENTS

The fly-by summary is OpenAI and Google announced advancements in personal, voice-activated AI companions that aim to provide more natural and personalized voice interactions, mimicking real conversations and adapting to user preferences and habits.

Introducing GPT-4o / OpenAI, YouTube (26 minute video)

  • try it via links on OpenAIā€™s blog post, Hello GPT-4o / OpenAI (3 minute read)

Google Keynote (Google I/O ā€˜24) / Google, YouTube (1 hour, 52 minute video)

  • highlight: Demis Hassabis introducing a low-latency demo with a phone and AR glasses (caveat emptor with Google ā€œdemosā€); starts at 26:10 (direct link)

some analyses:

Improvements to data analysis in ChatGPT ā€” Interact with tables and charts and add files directly from Google Drive and Microsoft OneDrive. / OpenAI (4 minute read)

Google is redesigning its search engine ā€” and itā€™s AI all the way down ā€” From ā€˜AI Overviewsā€™ to automatic categorization, Google is bringing AI to practically every part of the search process. / The Verge (6 minute read)

What GPT-4o illustrates about AI Regulation ā€” The important difference between regulating technology use and regulating conduct / Hyperdimensional, Substack (sorry) (6 minute read)

on education:

Googleā€™s new LearnLM AI model focuses on education ā€” LearnLM is already integrated into Google products like Android and YouTube / The Verge (2 minute read)

Math problems with GPT-4o with Sal Khan / OpenAI, YouTube (3 minute video)

 

šŸ“» QUOTE OF THE WEEK

Being a writer is the best way I know how to get paid for being insane.

Fredrik Backman (source)

 

šŸ—ļø FOUNDATIONS & CULTURE

Generative AI Is Totally Shameless. I Want to Be It ā€” The best thing about brain-melting software like ChatGPT? It doesnā€™t feel remorse. / Wired (5 minute read)

AI and the cost of failure / Tom Loosemore, Wordpress (3 minute read)

Start by researching the aggregate cost of failure. Estimate the cost of all the different ways you might fail people, from the trivial to the catastrophic (aka failure modes). And estimate the cost your failure imposes on your users; not just on your organisation.

  • related, How many people realise the answer is wrong?...At what cost to them and to you? / Sarah Gold, Linkedin (1 minute read)

    Using context specific UI to surface confidence levels in a suggestion made by an AI customer support system. Then, weā€™re wrapping this in a service that supports people to understand the implications of different levels of confidence. And seek re-dress if the system is less than fully confident.

 

šŸŽ“ EDUCATION

Ava enhances the counseling process with its expert-driven content, sourced from more than 277 top specialists across 110 topics, ensuring advice is both personalized and aligned with current academic standards. It offers comprehensive, personalized road maps that detail month-by-month planning tailored to each student's academic progress, extracurricular activities, and personal interests.

In Wisconsin, Chippewa Valley Technical College is keeping an eye on what the AI Incubator Network is producing, though it is not part of the network itself. In response to changing work requirements, the college is overhauling some of its existing programs, including the administrative professional associateā€™s degree program.

GeoGebra is more than a set of free tools to do math. Itā€™s a platform to connect enthusiastic teachers and students and offer them a new way to explore and learn about math.

A quarter of public K-12 teachers say using AI tools in K-12 education does more harm than good. About a third (32%) say there is about an equal mix of benefit and harm, while only 6% say it does more good than harm. Another 35% say they arenā€™t sure.

 

šŸ“Š DATA & TECHNOLOGY

How should you adopt LLMs? / Irrational Exuberance (12 minute read)

Behaviors do vary across models, but itā€™s also true that behavior of existing models varies over time (e.g. GPT 3.5 allegedly got ā€œlazierā€ over time), which means the overhead of dealing with model differences is unavoidable even if you only adopt one. Altogether, vendor lock-in for models is very low from a technical perspective, although there is some lock-in created by regulatory overhead, for example itā€™s potentially painful to update your Data Processing Agreement multiple times, combined with the notification delay, to support multiple model vendors.

The Limits of Data ā€” Data is powerful because itā€™s universal. The cost is context. / Issues in Science and Technology (19 minute read)

  • related (1), Protecting users with differentially private synthetic training data / Google Research (12 minute read)

  • related (2), Here's How U.S. Intelligence Will Buy And Use Your Data / Forever Wars (12 minute read)

    The data-purchasing framework issued by ODNI last week, which you can read here, creates rules for facilitating purchases, not hindering them, as much as they say that the ā€œprotection of privacy and civil libertiesā€ will be ā€œintegral considerations.ā€ Most important is what's not here. ODNI's nine ā€œgeneral principlesā€ don't require the intelligence agencies to purge any purchased dataā€”something Wyden urged when data acquisition surpasses privacy guidelines set by the Federal Trade Commission.

AnythingLLM ā€” The ultimate AI business intelligence tool. Any LLM, any document, full control, full privacy.

  • at capacity for hosted

  • self-host here: AnythingLLM: The all-in-one Desktop & Docker AI application with full RAG and AI Agent capabilities. / Mintplex Labs, Github

  • possibly related, LLMWhisperer

    LLMWhisperer is a technology that presents data from complex documents (different designs and formats) to LLMs in a way that they can best understand.

    • requires sign-up

    • adequate free tier for testing; obviously, donā€™t give it bank statements or anything sensitive

 

šŸŽ‰ FUN and/or PRACTICAL THINGS

Microsoft Places includes a dedicated location plan section where you can set and share the days youā€™ll use the office and view which days your co-workers are proposing to head in. Managers can set up priority days for in-office plans, so if thereā€™s an important event or a team day, everyone knows about it.

  • counterpoint (1), Hybrid Workplaces Are Still a Headache / The Walrus (13 minute read)

    Another issue around hybrid work schedules, particularly schedules that arenā€™t the same for all workers, is their potential for inadvertently creating a status divide based on the amount of time a worker spends at the office. As Charlie Warzel and Anne Helen Petersen have noted in Out of Office, hybrid work schedules threaten to deepen already existing divides between those currently in favour with the boss and those less so. Say Warzel and Petersen, ā€œSingle parents, workers with elderly family members, disabled employees, and those who simply donā€™t want to live in proximity to the office risk being overshadowed by those who come in every day. And even if a manager is careful, a recency and proximity bias might emerge.ā€

  • counterpoint (2), Ordered back to the office, top tech talent left instead, study finds ā€” In the months following return-to-office mandates, an increased number of senior employees departed Apple, Microsoft and SpaceX, often to work for competitors. / Washington Post (7 minute read)

  • counterpoint (3), Job Flexibility, Job Security, and Mental Health Among US Working Adults / Journal of the American Medical Association (31 minute read)

    Findings: In this cross-sectional study of 18ā€Æ144 US adults who were employed, greater job flexibility was significantly associated with reduced odds of experiencing serious psychological distress and experiencing anxiety. Greater job security was significantly associated with reduced odds of experiencing serious psychological distress and experiencing anxiety.

For Conversations You Dread, Try a Chatbot ā€” Role-playing with an AI conversationalist can prepare you to handle difficult subjects with family, friends and colleagues / Wall Street Journal (7 minute read)

How Airlines Are Using AI to Make Flying Easier ā€” Airlines are using artificial intelligence to save fuel, keep customers informed and hold connecting flights for delayed passengers. Hereā€™s what to expect. / New York Times (6 minute read)

via nic, 101 real-world gen AI use cases from the world's leading organizations / Google Cloud Blog (21 minute read; 3 minute skim)

 

šŸ§æ AI-ADJACENT

Smry ā€” AI Summarizer and Free Paywall Remover

  • not as effective as archive.today or 12ft.io in our testing, but still nice to see minor innovation in this space

  • summarizer worked one time out of ten; your mileage may vary

 

ā‹„