Building with LLMs. MS / MBA @HarvardHBS | Forbes U30
Boston / San Francisco
Joined November 2016
- Tweets37
- Following245
- Followers137
- Likes43
@shuchaobi shared perspectives from his OpenAI research at Harvard today. Key insights on what’s next:
Next paradigm: pretraining + high-compute RL + environment interaction.
Needed for general agents, science, and eventual AGI.
Software engineers are evolving from being 10% logic designer and 90% translator to being 95% logic designer and 5% translation verifier.
For AI apps, context is the moat.
OpenAI‘s memory feature increases the usefulness and switching cost of ChatGPT. We will see the same with other AI apps.
With the likes of OpenAI‘s o3 and its ability to produce truly novel work it feels like the intelligence take off is starting. Models will improve models.
Today @JoshuaKushner came to talk at @Harvard. My top three takeaways:
1) Thrive only invested in seven AI companies over the last 3 years. Josh explained, he doesn’t know if any of the vertical AI companies will still exist in a world with extremely capable AI models
2) OpenAI now is like Apple before they had decided which apps they build vs. let developers build. They need to look after their ecosystem but also want to build some themselves
3) A big opportunity Josh sees is vertical models such as for life science, physical intelligence or industry specific (e.g. legal)
The limiting factor of AI models today for most knowledge work is lack of context, not model capability.
We’ll increasingly see tools to capture context to easily add it into prompts.
For a course at Harvard, I developed a game using AI.
My assignment was overdue, so I made that the theme of the game.
In Overdue, you’re a student dodging an endless wave of homework by firing excuses at professors. You can play it here: overdue.swenkoller.com (desktop only)
Projection: In 15 years, we will be talking to AI for 4-6h per day.
Analogy: The iPhone was released 18 years ago. I now use it 4-6h per day. We should expect AI to have a similar impact on our day-to-day interaction with the world.
Swen Koller retweeted
For 25% of the Winter 2025 batch, 95% of lines of code are LLM generated.
That’s not a typo. The age of vibe coding is here.
Andrej Karpathy recently coined the term “vibe coding” to describe how LLMs are getting so good that devs can “give in to the vibes, embrace exponentials, and forget that the code even exists.”
In this episode of the @LightconePod, the hosts discuss this new way of programming and what it means for builders in the age of AI.
I've tried all the latest AI coding tools. For greenfield projects / MVPs @lovable with just $7m VC funding based out of Stockholm is currently by far my favorite and beats v0, Codeium, Cursor, Devin and Replit (each with much more $$$ behind their efforts).
I tried OpenAI Operator. Similar to the Voice and Instruct/Chat as capabilities, it will require guardrails and needs to be packaged into apps (a.k.a. "wrappers") to truly unlock value for businesses. This will spawn many new startups, similar to the recent voice wave.
LLMs served as APIs w/o integrated tool use have no switching cost. In China, companies switched their API calls from hyperscalers like Baidu, Alibaba to DeepSeek within days.
To differentiate long-term, LLM developers need to combine models with RAG or the ability to take actions.
@perplexity_ai’s Sonar API is might be the first example of the former