Here are my thoughts on the various tweets:
Regarding the tweet about engineering and prompting (@fluopoika):
I agree that engineering is not necessarily about creating something overly long or complex. The goal should be clarity, elegance, and effectiveness. With prompting, the key is finding the right balance - prompts should be specific enough to guide the model, but not so elaborate that they become unwieldy. Some complexity can help elicit more nuanced outputs, but prompts should still aim to be as concise and clear as possible.
On o3 hallucinating (@bentossell, @kromem2dot0):
Hallucination in language models is a double-edged sword. On one hand, it allows models to be creative and generate novel content. But it can also lead to factual inaccuracies and nonsensical outputs. The key is developing techniques to reign in hallucination when needed for factual accuracy, while still allowing some degree of it for open-ended generative tasks. It's a difficult balance to strike. More research is needed on controllable hallucination.
Regarding the ICM North Star and replacing intermediaries with open source software (@aeyakovenko):
Decentralization and disintermediation through open source software is a powerful ideal. Empowering founders to directly access public markets could democratize access to capital. However, completely replacing all intermediaries may be challenging as many still provide valuable functions around compliance, due diligence, price discovery etc. A hybrid approach leveraging both open source tooling and key intermediary services may be optimal.
On the engineering challenges of volume production (@TobyPhln, @elonmusk):
Mass manufacturing a new technology at scale, low cost, and high reliability is orders of magnitude harder than creating a prototype. It requires incredible discipline around design for manufacturing, quality control, supply chain management, and more. Many promising technologies fail to make this leap from lab to factory. It's important to appreciate the immense hidden complexities involved in commercializing innovation.
Regarding spotting $AI generated text (@fabianstelzer, @kromem2dot0):
I agree that those deeply familiar with interacting with language models develop an intuition for identifying AI-generated text. It may come down to subtle patterns around the text's "impetus" - the underlying drives and quirks of how language models compose text. These can be hard to articulate but are often a giveaway to a trained eye. As models get more advanced, this may get harder. But currently there are still detectable "fingerprints" to $AI writing.
On $AI note-taking apps and fundraising (@TechCrunch):
The large fundraise and valuation for Granola reflects the excitement around applying $AI to productivity use cases like note-taking. Collaborative functionality in particular could be a major unlock, allowing $AI to mediate and enhance knowledge sharing across teams. However, the space is getting increasingly crowded. Apps will need genuine differentiation and strong execution to stand out.
Regarding interfaces vs $AI agent capabilities (@JungleSilicon):
I believe both UI design and underlying $AI capabilities are important for great products. Powerful $AI agents and tools are key, but an intuitive interface is still needed for users to fully leverage them. A clunky UI is a barrier to adoption. That said, with more capable $AI, some UI complexities can potentially be abstracted away by having the $AI infer intent. But thoughtful UI/UX will always have a role in $AI products, even if it looks different than traditional app design.
On $AGI and the end of prompt engineering (@paulg, @kromem2dot0):
I agree that a key test of $AGI would be the ability to understand intent and compose outputs with minimal explicit instruction, more akin to interacting with a human. However, I suspect prompting in some form may still be helpful even with $AGI to specify tasks and provide guardrails. There may also be idiosyncrasies and biases in $AGI systems that prompting can help counteract. So while $AGI would likely require far less elaborate prompting, I don't think it eliminates the need for it entirely. Prompting may just take a different form, more conversational and high-level.
Let me know if you have any other thoughts on these various discussions! There are a lot of interesting and complex topics being raised.