weekend ai reads for 2024-06-07

June 07, 2024

📰 ABOVE THE FOLD: ON BUILDING WITH AI

Home-Cooked Software and Barefoot Developers — The emerging golden age of home-cooked software, barefoot developers, and why the local-first community should help build it / Maggie Appleton (24 minute read)

  • summary of a talk, with slides

  • hopeful point-of-view that posits since the barrier to website development is basically zero (for a certain subset of the global population, that is), there will create more interesting and unique websites for everyone

So I think you should care about this because the local-first movement and the local, home-cooked software vision are distinct but philosophically aligned.

They’re built on the same foundational values: that users should have agency and ownership over their data and software.

At the moment this community is focused on solving hard technical problems, but you should keep an eye on what’s developing around you as parts of the software-making process rapidly become more accessible and democratized.

LLM development-prod skew can be categorized into two types: structural and content-based. Structural skew includes issues like formatting discrepancies, such as differences between a JSON dictionary with a list-type value and a JSON list, inconsistent casing, and errors like typos or sentence fragments. These errors can lead to unpredictable model performance because different LLMs are trained on specific data formats, and prompts can be highly sensitive to minor changes. Content-based or “semantic” skew refers to differences in the meaning or context of the data.

  • long reads but a lot of useful and practical insights for anyone building anything with LLMs

  • Part III is coming

How to fine-tune Phi-3 Vision on a custom dataset / W&B Fully Connected, Weights & Biases (13 minute read)

  • related: Open Engineer — Open Engineer is an entirely open and free resource for learning AI-based technical skills, designed to be helpful for beginners and experts alike.

 

📻 QUOTE OF THE WEEK

We’re building an AI tool with OpenAI in which you’re flying your little robot and now you're next to a messenger ribonucleic acid, which is what the COVID vaccine was made of.

You’re going to probably become a master biologist, just by asking this machine.

Michael Crow, President, Arizona State University (source)

 

🏗️ FOUNDATIONS & CULTURE

AI saving humans from the emotional toll of monitoring hate speech / Waterloo News, University of Waterloo (4 minute read)

Unlike previous efforts, the Waterloo team built and trained their model on a dataset consisting not only of isolated hateful comments but also the context for those comments.

SPE will look to “produce both films for theaters and television in a more efficient way, using AI primarily,” he said.

Don’t Believe the AI Hype / Daron Acemoglu, Project Syndicate (7 minute read)

If you listen to tech industry leaders, business-sector forecasters, and much of the media, you may believe that recent advances in generative AI will soon bring extraordinary productivity benefits, revolutionizing life as we know it. Yet neither economic theory nor the data support such exuberant forecasts.

 

🎓 EDUCATION

Does Edtech Need A Hard Reset? — New York Times Opinion columnist Jessica Grose says we need to “Get Tech Out of the Classroom Before It’s Too Late.” Here’s our take. / Whiteboard Advisors (5 minute read)

On Building AI Models for Education / AI Education, Substack (sorry) (20 minute read)

  • the first takeaway:

Creating an effective AI tutor is tricky because we don't have great ways to measure if the AI is really helping students learn better. We need to develop better, standardized ways to measure how well AI tutors teach. There’s still a lot to do!

Google, AI Announcements, and the Future of Learning — For real transformation, we'll need learning-focused solutions instead of tech-focused ones / On EdTech (11 minute read)

But even beyond the misreading Bloom problem, thus far I am unconvinced that the kinds of tutoring currently offered via AI matches the concept of watching a student’s thought processes and identifying the core issues they aren’t understanding. Instead, AI tutoring today seems to consist of breaking down problems into component parts and explaining the components. This is no doubt helpful, but it is not tutoring in the true sense of the word.

 

📊 DATA & TECHNOLOGY

Tempus AI, Inc., Form S-1 / Securities and Exchange Commission

  • form Tempus AI filed in anticipation of its initial public offering

  • deep insight into the workings of a thriving AI company, especially the role data play in their competitive advantage

  • focus on the prospectus summary (44 minute read)

  • and the risk factors (229 minute read, but you can skip all the geo-political risks so much quicker)

Our ability to maintain, expand and monetize our datasets are subject to a number of factors, many of which are outside of our control. With respect to data included in our Data and AI Applications products, we rely on a combination of the statutory rights available to us as a HIPAA covered entity and as a HIPAA business associate. As a HIPAA covered entity, we utilize data generated through our provision of Genomic tests.

Scale's SEAL Leaderboards / Scale.ai Blog (7 minute read)

The SEAL Leaderboards are a set of LLM model rankings across a number of popular public models, based upon curated private datasets that can’t be gamed, all funded and developed by Scale.

Benchmarking foundation models for time series / nixtla, Github (12 minute read)

Foundation models for time series outperform alternatives and are ready to be tested in production. TimeGPT-1 is (so far) the most accurate and fastest model but TimesFM from Google comes very close. Some models are still outperformed by classical alternatives.

 

🎉 FUN and/or PRACTICAL THINGS

Zoom CEO Eric Yuan wants AI clones in meetings / The Verge (49 minute read)

Let’s say the team is waiting for the CEO to make a decision or maybe some meaningful conversation, my digital twin really can represent me and also can be part of the decision making process.

The conversation with the future self is generated in real-time via a large language model that has been personalized based on a pre-intervention survey assessing user future goals and personal qualities.

  • the app, Future You

  • also, this either promises 10-hour workweeks or portends greater economic inequities

How to Use Copilot In Word / Lifewire (5 minute read)

Music Lawyer AI — Providing artists with legal clarity by identifying potential issues with recording agreements

  • exclusively for the one reader for whom this may be relevant

Keeper — Optimize your taxes with the power of AI

  • using a famously bad math tool for a high-stakes problem that could land you in jail if you get it wrong seems fraught with some risk

  • they claim everything is checked by “a network of tax professionals”

 

🧿 AI-ADJACENT

Cara — Artist Social & Portfolio Platform

  • an artist-centric site that aims to keep AI-generated out

and …

The New Generation of Online Culture Curators / The New Yorker (9 minute read)

Perhaps the best way to think of these guides is as curators; like a museum curator pulling works together for an exhibition, they organize the avalanche of online content into something coherent and comprehensible, restoring missing context and building narratives. They highlight valuable things that we less-expert Internet surfers are likely to miss.

 

⋄