[NYC AI Event Pick] | April 2025 & What can AI do today, really?
A lot more then you think... but also not so much.
💡 Editor’s Note: There has been lots of hypes going into this year. AI Agents, world foundation models (yes, the wall-E on GTC), deep research, no-code platform like v0.dev and lovable that make apps for you in minutes. Piling on this, the new image models from Gemini & ChatGPT adds an additional layer of surrealism. All these hype aside, what we can really, reliably do today with AI? We will dive into this topic in our Builder's Corner and show case some actual uses cases that have been quite useful - for me at least : D
P.S. You can find table of contents (TOC) on the left.
🎉 Upcoming AI Events - April 2025
🌟 Featured Events
[IRL] Manus AI Fireside Chat | Mar 31 | RSVP Exclusive opportunity to meet Tao Cheung, Manus co-founder.
[Application] Mini-Hackathon v1 | Apr 5 | RSVP Put AI productivity to the test! Go from idea to MVP in just one day with AI Tinkerers NYC.
[Deep Dive] Cornell Tech Media Tech Summit 2025 | Apr 25 | RSVP Inaugural summit bringing together creators, technologists, researchers, and entrepreneurs to explore AI and XR's impact on media. Four powerful panels cover AI content creation, immersive technologies, trust in digital media, and entrepreneurship opportunities.
📅 Key Events This Month
IRL
NYC GenAI Companies Showcase | Apr 9 | Apply Meet founders building breakthrough AI innovations, hosted by The AI Furnace.
Agents are Taking Over NYC with Kamiwaza | Apr 16 | RSVP Evening of demos and networking focused on enterprise-grade AI agents and open-source agent infrastructure.
Application
Climate Adaptation in the AI Era | Apr 1 | RSVP Brown University hosts Roberta De Vito discussing AI/ML applications for climate resilience and addressing technology's environmental impact.
Unpacking Ethics in AI with Margaret Mitchell | Apr 2 | RSVP Practical insights on AI ethics through real-world examples from recent generative AI developments.
Scaling AI with Next Generation Energy Solutions | Apr 3 | RSVP NYU explores how AI growth is reshaping energy consumption and how industries can innovate to meet scaling demands.
AI Ignition Engine: Business Value Assessment | Apr 7 | RSVP Strategic reality check on where AI delivers value and how to avoid costly implementation mistakes.
Learn How to Build A Custom Chatbot | Apr 8 | RSVP Hands-on session exploring chatbot applications for exam prep, career development, and business use cases.
Deep Dive
The US-China AI Contest | Apr 2 | RSVP Dr. Michael Mazarr (RAND) discusses political and technical dimensions of US-China AI competition.
AI in Production: Video Generation with Modal Labs | Apr 2 | RSVP Hands-on workshop on creating and deploying custom generative video models using open-source foundation models.
AI: Advancing Foundational Biology | Apr 8 | RSVP MIT/Whitehead Institute symposium exploring AI/ML transformations in biological research with experts from leading institutions.
Using AI in Qualitative Research | Apr 11 | RSVP NVivo AI demonstration and expert panel on promises and pitfalls of AI in qualitative research methods.
Builder's Corner - AI for productivity... How?
TL;DR. This article showcases productivity examples with reasoning models (basically, models that think) and where it falls short.
Now, I must confess I am a pretty heavy AI user. So far this month I have sent 5.5 Million tokens to o1pro, and used 665 premium requests. So my why view towards whether AI is useful or not could be quite biased : D
So far this year, I have been using AI for creating toys and arts for my toddler, building and launching applications (web & mobile), gathering information and managing those knowledge, organizing my activities and emails, practicing my public speaking skills, refining ideas, managing the NGO and writing this newsletter even.
Did AI save me time and money on this efforts? Yes. But it doesn't always meet my bar for quality...
1. First off, which model?
There are many different types of model today for different modalities: text, audio, image, video. Today, we focus on the text models. Specifically reasoning and non-reasoning models. What are reasoning models?
Now this topic can be expanded to a much longer article on reasoning in models and humans... which I won't digger further today. However, I would refer curious readers to an article on human reasoning and a collection of articles on LLM reasoning.
Now, simply put, reasoning models they are model that 'thinks'. Like o-series models, gemini thinking (2.5), claude 3.7 extended and deepseek r1. So when should I use reasoning models?
OpenAI has a lengthy article about which model to choose and when:
Speed and cost → GPT models are faster and tend to cost less
Executing well defined tasks → GPT models handle explicitly defined tasks well
Accuracy and reliability → o-series models are reliable decision makers
Complex problem-solving → o-series models work through ambiguity and complexity
Now... if you have no concern on cost, I would definitely recommend using reasoning models for all complex tasks, and 4.5 for tasks focused on contents. Personally I have been using primarily o-series models since it's release and other reasoning models. This means:
Try o3-mini-high and o1 (or o1 pro if you have it)
Try Claude 3.7 Extended
Try Gemini 2.5 & 2.0 Thinking
Last but not least: Don't use DeepSeek R1. According the most recent hallucination benchmark, deepseek-r1 has a whopping ~14% hallucination rate. In contrast, the models above have average hallucination between 1%-2%.
2. Productivity Gains
My productivity gains from AI comes in two forms: AI augmented workflows and AI created workflows. The former refers to when AI is directly tied to the core logic of my task, while the latter is when AI is used to create tools that handle the core logic of my task - this are what AI alone cannot do.
Now, I will examine the efficiency of AI through the creation of the newsletter. I hope it won’t be too surprising when I mention that this newsletter was created with the help of AI. Can you guess which part? : D
There are literally thousands of AI events happening every single month across the nation - the big question is which ones are worth going? Now, this varies from people to people, but there are at least two common criteria: 1. Relevant 2. High Quality.
The initial step shown in Figure 1. was usually done by hand - selecting the events that meets the criteria. For this month, I have tried two Agentic AI tools to help, both failed to different degrees.
OpenAI Operator: Found 1 event and stopped. (Example)
Manus: Found 22 events, 3 were usable. (Example)
So for now, step 1 will continue to be manual. However, step 2-4 could be reliably delegated to AI!
I have done all of these steps manually, before reasoning models were available, so I can tell, definitively, how much time AI has saved me : D
2.1 AI Augmented Workflows (Time Saved: 3 Hr/Month)
For the flow, initially, the event information are collected into a google sheet, this creates a large csv file with tons of information. Previously I had to clean up this information manually (as you can see, non-reasoning models don’t quite work). However, with reasoning model, I was able to get much reliable results.
Specifically, AI augmented workflows helped me through step 2-3 and 4B.
The top left of figure 2 is the raw event information collected manually. The top right, is the formatted CSV files using ChatGPT o1-pro. The bottom left is the output copied back to the google sheet. Prompt is simple and direct as below.
Note: For the intricacies of prompt engineering, it’s components I will refer readers to Chapter 5 of AI Engineering by Chip.
Prompt:
<system> output in CSV format </system>
<task>Reformat the events information I have from @Events into the csv header format: Title Cost StartTime EndTime EventType Location Link Description Affiliation</task>
<Events>…..</Events> <Example>…</Example>
Now, I wouldn’t be a geek if I don’t compare this methodically. So here we are:
Non Reasoning: 4O Result, Gemini 2.0, 3.7 Sonnet, Deepseek V3
Reasoning: O1 Pro, O3 Mini H, Gemini 2.0 (T), 3.7 Sonnet (T), Deepseek R1
As you can see in the results, non reasoning model all messed up the start/end time, which should be in the following format “04/08/2025 14:00:00”. Reasoning models generally handled this better (O3 Mini H thought really hard but failed.. could be a bug though) .
Next up, the events section above was created using Claude 3.7 as part of step 4B. With the following prompt in one go
Prompt:
Help me draft an event newsletter (in markdown) based on @events . I need very concise, to the point event description, with links to RSVP for the events. Each event should take no more than 2 lines. Example row: Name, Date, Link, Target Audience, Content
<events>…</events>
With reasoning model, it is much less of a concern that the model will omit key information - and we can be very demanding with our tasks.
2.2 AI Created Workflows (Timed Saved: 1.5 Hr/Month)
Now, there are tasks that AI cannot do. Specifically, in the context of this newsletter, it’s creating the events on Google calendar (Bottom right of Figure 2).
The entire flow is quite simple actually. Without AI, I would have set up automation tools like Zapier or Make to create a workflow like below. The downside? I need to PAY for it!
Now, of course, I could pay for it, but for such a simple task like this one. I have absolutely 0 interest paying ~$10-$20 a month for it. So, I opted to do it myself and get the function for free.
The workflow is quite simple, as show in Figure 4. You can first write a prompt like the one below to generate the initial AppScript code, and then iterate on it.
Prompt:
I have a google sheet with the following columns [Title Cost StartTime EndTime EventType Location Link Description Affiliation]. I want to create a google calendar event for calendar “$myemail@gmail.com”, the event title should be based on Title, the link should be Link, the location should be Location. Now, the description should be a summary of the [Cost EventType Description Affiliation], and I want the summary to be written using OpenAI api. Prompt user for api key.
<task>Write a AppScript to finish the task </task>
Check out the initial code from the above prompt. This usually takes just 2-3 iterations to get it down. Since reasoning models are much much better at leveraging contexts, it’s best you throw in large contexts and then reference them in the prompts - the model will go and figure it out.
In short, using AI to create this workflow have saved me 1.5 hours/month if I were doing this manually, or ~$10-20/month if I pay for automation tools. Doesn’t sound like a lot right? But this is just one simple use case and you could create more!
Now, this is NOT how you should be coding for even moderately complex projects, where the issue becomes context management, which is a common issue when coding using cursor and other tools (Repomix is a good tool to bundle all context in one go btw).
Remember: With AI, you no longer need to pay a fortune to perform all of the simple tasks - email summary, news summary, event creation, calendar management etc. Just DIY - for free. If you are interested, I can cover these examples in future posts : D
3. Parting Words
In this post we have walked through how to efficiently use reasoning model through a few examples. This is only the beginning of what we can do with AI today.
One particular issue I mentioned in section 2, is that
AI tools are inherently probabilistic and quality is a concern
This still holds true for all of the tasks I have described above. I still need to manually verify the event links created, the time, content description for all the events created. I need to verify if the code generated is correct, and iterate through testing.
To learn more, one person I would definitely follow is Andrej Karpathy (ex-OpenAI, now Eureka Labs). This video show quite a few use cases that could hopefully spark some ideas.
Stay tuned—next time, I’ll reveal the specific tech stack that powers these projects, including how I integrate tools like Cursor, v0.dev, and beyond!
Next up in community
We just wrapped out Educational event with Newtown Public High school (Thanks to Google). In Q2, we will host several more events, so stay tuned! !
