cooking with ai #4 - ai for blogging, slide generation (failure), experimental releases like notebookLM, and zoom's ai avatar

welcome to the fourth episode of cooking with ai, where we explore practical experiments with ai tools and discuss emerging trends in the field. this week, we dive into our test kitchen experiments, examine no-code ai tools, and explore the future of ai interactions.

in this episode, we cover:

1. ai-powered blog generation by will
2. test kitchen failure: slide generation
3. archetype's newton ai, a foundation model for physics
4. the age of experimental releases
5. zoom releases ai avatars

to accompany episode 4, this is the list for all the topics and resources discussed. we're exploring ways to make the content shared more skimmable and accessible for future reference.

enjoy!

1. ai-powered blog generation by will

our first experiment focused on using ai to generate blog posts from podcast transcripts. will's custom blog engine, designed to process and transform conversations into coherent narratives, revealed interesting insights about content generation.

the experiment highlighted a crucial learning: different types of content require different processing approaches. while the engine worked well for a single-topic, focused content, it struggled with the meandering nature of our conversations.

this challenge pointed to a broader insight about ai content generation: the need for specialized tools for different content types and objectives.

?? here's will's blog.

2. test kitchen failure: slide generation

in our quest to bridge the gap between research papers and visual understanding, we tested several no-code ai tools to help summarize and make slides.

they did not work as we had hoped. this is a process that requires multiple steps to do well.

and in the spirit of cooking up things in the ai test kitchen, we wanted to share an honest update, which includes the failures.

we started with archetype's new research paper on their newton ai, a foundation model for a transformer based model to understand real world physics. for each app, we shared the research paper and also a notebooklm summary of it to see what would happen.

trying out gamma app

• produced basic slides with shallow content
• better suited for simple presentations
• worked better with a pre-processed summary than with raw paper

?? full gamma here.

testing prezi ai

• generated strange graphics
• information extraction similar to gamma, ie. limited
• complex transitions hindered understanding
• ?? full prezi here.

mylens ai

• created hierarchical mind maps
• over-generalized content
• ?? lost specific paper details

?? full mind map here.

napkin ai

• showed promise with diagram generation
• better at technical visualization
• more suitable for collaborative editing

?? full napkin ai here.

final thoughts for slide generation

while each editor had the ability to generate some sort of visual automatically and the ability to share via weblink, these test kitchen experiments were a failure.

to be able to effectively translate complex information into helpful slides, these tools either oversimplify the content and/or have a disconnect with the visuals. this seems to require more of a specialized workflow that separates steps when parsing the paper and making the slides.

this highlighted the importance of understanding why use a tool in the first place to better break down the end goal and match to workflow.

• if the goal is summarize for beginning to understand the content, then a summarization prompt works better in chat window with an llm than these tools.
• if the goal is to make custom diagrams for visuals for slides, then focus on new visual generation.
• if the goal is a complete presentation with slides that could be presented right away, then it will require a specific tool that can handle research papers as links or pdfs.

3. newton ai by archetype ai

so while the tools did not work to make a shareable slide deck for this release, it is a fun one to cover.

archetype ai trained a foundation model to take in data from real world events, like sensor information, to be able to predict outcomes from new situations, like the movement of the weight and spring system above.

while open ai's chat gpt model is good for natural language prediction, there is now a search to other general pretrained transformer (gpt) models to work on other kinds of data, like sensor information.

the goal for newton ai is to help with physical system prediction like:

• city electrical demand
• daily temperature
• temperature in electrical transformers

the benefit of using a transformer architecture is similar to chatGPT: one model, many use cases, less training data and compute power for new use cases, and fine tuning with new data.

?? full release here from archetype ai.

4. the age of experimental releases

google's approach to ai product releases through labs.google represents a significant shift in how major tech companies are approaching innovation. rather than having large, polished releases of core products, there are incremental releases and batch tests for new products. this helps:

• lower the cost to release these new products
• small teams work faster without traditional approval processes
• speed up the iteration cycle with direct feedback from customers

this is also the first time google is taking a technology first approach to products versus focusing on the business and use case first, like what alphabet and their x moonshot lab focuses on. it's impressive to see the results, especially within a large company, where these kind of product releases can be hard.

some takeaways from notebooklm team:

• there still needs to be a secret sauce for product value. for this product, it's their new audio model for smooth podcast quality
• with the bar for deployment being lowered by access to foundation models, the expectations are higher for result. one engineer on the team focused on ensuring great podcast scripting.
• delight matters. this product feels like magic, similar to suno ai, where one input and one button produces a whole podcast, where users are unsure what they might get.

alongside google labs release, we're also seeing this with open ai's o1 preview release, which enables paid users to try new model before full roll out.

where this starts to go too far on early release is something like apple intelligence. that was the first set of features shared at a wwdc event that was not ready to be shipped. iphone 16s have started shipping and while they have the capability for apple intelligence, the features are not yet available.

access is an important part to experimental releases to maintain trust with users and bring them into the process of development.

5. ai avatars: the next frontier of digital interaction

ai avatars are nothing new. hey gen and synthesia's platforms enable avatars for video generation.

zoom's recent introduction of ai avatars has sparked discussions about the future of digital interactions. would your ai avatar talk to another ai avatar for meetings, where you only read the transcript? for power dynamics, would your boss be able to send their ai avatar?

while the immediate reaction has been mixed, our take is that if people don't want it, then it won't be adopted.

rather than think about in current forms, we discussed the ways in which it could be interesting for recruiting and interviews:

• enabling recurring interviews with feedback scores, rather than one and done evaluations
• connected to interactive prompts and multimodal input, like smart whiteboarding to be a true co-pilot with new functionality

looking ahead, the success of ai avatars might depend less on their ability to replicate human interaction and more on finding unique use cases where they provide distinct value.

looking forward

as we continue cooking up our experiments, one thing becomes clear: we're in an era of rapid iteration and experimentation. the key to success lies not in perfect execution but in thoughtful experimentation and learning from both successes and failures.

what's cooking next week? stay tuned as we continue to explore new tools, share our experiments, and discuss the evolving landscape of ai technology.

thanks for making it this far. if you have topics you want to see in future episodes, or ideas for what you may enjoy, we want to hear from you! comment below or send me a dm.