Connor White
@bconnorwhite
Followers
287
Following
5K
Media
39
Statuses
1K
Retired Software Eng. / Full-time Tinkerer
Austin, TX
Joined August 2015
@autosseyai One of the most powerful elements in Prxmpt is the <priority> element, which is sort of like a CSS Media Query for text. It enables fine-grained control over which sections of your prompt should be prioritized once you run out of tokens:.
0
0
1
Super excited to release Prxmpt!. From using this while working on @autosseyai I can say that writing prompts with JSX is the future. Prxmpt makes it far easier to format complicated prompts. If you use TypeScript + OpenAI, give it a try and let me know what you think.
Introducing Prxmpt: a JSX framework for composing prompts. Prxmpt is like "TailwindCSS for Prompt Engineering" -. A set of utility elements that make formatting complex prompts super simple. No more convoluted template literals or nested ternaries.
2
0
6
Seeing self driving cars everywhere still blows my mind. 50k years ago there were 4 species of hominids and everyone lived outside.
0
0
1
With all the talk of Gemini, Google is finally rolling out a 32k context PaLM-2 models:. text-bison-32k.code-bison-32k.chat-bison-32k.codechat-bison-32k. Even an embedding model with longer context:. code-gecko-32k.
0
0
4
AIs become way more capable when you just give them good tools. Humans are the same way.
0
0
1
Dyson spheres have no moat, they're just a wrapper around the sun.
0
0
2
Seeing completely empty Cruise cars driving around Austin is wild, can't wait to try it out.
0
0
2
I can't wait to be able to whiteboard with multi-modal GPT-4.
0
0
6
"Prompt Engineering" is Software 3.0. But Software 3.0 is not the final paradigm:
0
2
3
Methods to improve LLM reliability:. 1. Make the task easy (split into smaller steps).2. Make the answer obvious (improve context quality).3. Provide multiple opportunities to succeed (check your work). Mistakes are more expensive than more context tokens.
0
0
3
Data formats that can be parsed by an LLM are just much easier to build tools around vs. APIs, or proprietary data formats. Companies will integrate AI into their own tools, but it will be harder to build systems that work across multiple applications. Where this could change.
0
0
0
3. CGI/CAD. Jobs working in 2d/3d visual environments (design, 3d modeling, etc.) will be relatively untouched until multi-modal systems are both cheap & extremely capable.
1
0
0
2. GUI. Jobs that involve many different GUI applications with a mix of text boxes, buttons, menus, etc. will see lots of "workflow enhancing" tools, but slower full automation than jobs or tasks which mostly involve directly manipulating text files.
1
0
0
1. CLI/WYSIWYG. Text based jobs (ex: programmer, writer) are a massive opportunity for highly capable systems. Niche products risk fast obsolescence as there is already a massive base model vs. product capability overhang here. Progress toward more advanced systems will be rapid.
1
0
0
CLI / GUI / CGI.(text-based / application-based / visual). As we get closer to Keyboard AGI (AI that is general over text modalities) the opportunity for products & resulting impact on different areas of knowledge work will vary based on how you interact with your work.
1
0
1
When you feel tired "just write a TODO list" is such a good hack. As soon as I start the list my curiosity takes over and my brain forgets to be tired.
0
0
1
When you accidentally send OpenAI your entire node_modules ðŸ˜. Fortunately it was GPT3.5, not GPT4.
0
0
2
Turing test, but the examiner writes issues and the AI writes pull requests.
0
0
0
ANAGI - Artificially Narrow AGI. A system which is able to perform general tasks if you give it enough time, but chooses to specialize by "hard coding" certain strategies to be more efficient in a specific domain. ANI -> AGI -> ANAGI. Generalize, then specialize.
1
0
2