Hey David, your most average fan here :) When you mentioned a 500% productivity increase that kinda floored me. Mine has been between 500-600% Writing a 513 page long cookbook from nothing, just hands on experience, took me 11*16 hours for the bulk writing, and 3*12 hours for all the little items to get ready for print on demand. That last chunk isn't all that automateable. So that's 212 hours with averages to a metric--npt the actual time, just a metric--to 2 1/4 pages per hour. Including artwork, testing, graphic design, iterative testing, yadda yadda. All of it. Basically 500+ pages in a work month. (I was pulling monumental hours though, that's how it ended up being a little shorter than two weeks) since Friday I've pulled 60 pages, a bunch of original art, my own little frame shop, food blogging, stabilized a bunch of my small business, yard work, and so much more. I don't think I've gone above, or have any realistic way of going above 600% increase. But even 50% would have been amazing. And a lot of it is because of You and Matt Woof.
As someone who started with a love for writing and reading, and transitioned to coding and DevOps, this video discusses the exact thing I love so much about generative AI: it allows you to play with English in different ways to get better results. I like to think of the way we talk to generative AI as "English as code".
Your ideas are ahead of the curve of what is current knowledge and what is discussed. Cognitive offload and finding the right tone to ask the LLM is something people haven’t really considered in the broad space. Just a very good collection of coherent ideas.
I have been using Gen A.I. for story writing for a few months now. The Polymorphic code slide would really push things to the limit. I agree with you, but that is alot to get my head around.
With a few hundred words of notes you can get an AI to generate an outline using the Hero's Journey template. A log line, protagonist, antagonist, love interest, some challenges would be enough. Maybe a synopsis. You may want to do some homework on process improvement techniques.
Hi David, thank you for this one! Did you look at h2oGpt AI? They claim to be an open source organization and want to make fine tuning LLM as simple as possible for the average joe! I think you will find it interesting and if so would be nice if you make a video on them. :)
33:46 any recommendations/links on the fastest path to leveraging tree of thought based context for agendas less boilerplate then a pre-existing and well understood context such as an occupational role? For example, personal multifaceted and standing agendas that are switched between throughout the day. Simplest examples being movement between personal and professional aspirations. Would be great if I could flip to my ____ contextual construct and then field that email that just came in or initiate a new research thread with all the standing context and research to date already in mind and ready to build. A well executed version of this idea could even generate things proactively. Planning to do this and will be building it if it’s not currently able to do what I have in mind. Any immediately viable solutions would be great - if you (Dave) or anyone else out there has any pro tips to pass on. ✌️
Will we have websites and apps that need internet access in the future? I wish i'll see a future where the internet reverts back to raw essential data streama and everyone will have their personal ai to display the data on whatever platform the user is looking at in the perfect way for that person
I've had great success with generating code. I think GPT4 (not 3.5) is good enough if you provide context and examples, and can judge the output. I would say 50%+ of the code I commit now was "written" by AI and just orchestrated by me. I really enjoy the workflow because using so much natural language helps me articulate the problem and think through it thoroughly. It's getting both the code and explanation out of my head and onto the screen.
What decisions are you making? It's funny because we seem to be doing the oppisite. I don't trust it to make architectural/important decisions for me etc, just write what I tell it to.
You've been keeping a steady pace of videos about one every day... Careful of burnout, a few ( two or three) a week is good till we pass the knee. Are you looking to hire some help to create content?
thats part of his 5x productivity;-) maybe he can create a pipeline to release them later so he can take time off whenever he wants, while keeping us happy;-)
agree on almost everything, but I seriously still don't think that we can generate code on demand (29:37). Not as it is. Code must be as deterministic as possible.
That was amazing David! Well researched and presented and mimics if not expounds and improves upon many of my thoughts and deployments in the space…great job once again and I tip my hat to you sir!🥳🦾🤖👏🏼
TL;DR Davids etimates remain very much in ballpark to my own experience with writing and generative image creation. Okay, because I really, really liked this video, I decided to break out the ol' timer, and see how much active creation, effort, formatting, image creation, link checking, hallucination checking, and so on it takes me to create the following: 1. Original detailed technical documentation with ChatGPT4, Plus. 2. All original images with Dall E via bing. Bing lightly watermarks it's images, so those had to be scrubbed. 3. Keep a document of most of the prompts I used. Some I just lightly edited a few times, so my actual number of prompts that I ran were ~3x more than the original prompts. I developed a couple of documents. I will link them to 'show how the sausage is made'. All GPL, CC, attribution. They allare around Information Security. And of course, if you are uncomfortable clicking a link here I understand. Heh, it is kinda what we infosec folk say NOT to do. (on LinkedIn I'm Jeremy Pickett westcoast, web home is www.japh.one, happy to link my bona fides). This is the document in question: docs.google.com/document/d/1BIAmeke9vZrl13Iio11xkkc8bnI64lT2/edit?usp=sharing&ouid=104982846581126963187&rtpof=true&sd=true Subsequent docs will be in replies to myself, so I clutter up main as little as possible. I grabbed the document/blog post prompt from a running spreadsheet of mine, created by (drum roll) ChatGPT. The prompt was, "Create a detailed specific blog post with four sections for the subject Threat Intelligence and Its Role in Strengthening Blue Teams". That created a response with ~2000 words. That is the seed. So there are actually six sections: Introduction, S I, S. II, SIII, S. IV, and Conclusion. Each section, while wrangling the responses in a Google doc with a fairly standard template I use a bunch (headers, footers, fonts), I took each section and ran a subsequent prompt. "Can you expand on Section 1, "Section I: Threat Intelligence - The First Line of Cyber Defense" of the document we named INT-1 into details and specific examples or use cases. Include links inline to the details in a natural manner. Use as many tokens, words, and sentences as you can. Explain it like a concise college grad student, do not guess. Link to webpages, urls, sources inline to the blog article." Then for each section do quite a few things: do the links work? 50% failure rate, so fix those. Is tone, language, fairuse/plagirism okay? Fix those. Is it correct? I've been in information security for 20 years, so I can judge that part really, really fast. Rewrite as needed, and make it sound less like a college freshman writing their first year essay while WAY TOO STONED for the time being. Get through all six. Then, for each section, try and find a piece of history, subject context, website summary, whatever that can compliment the larger section. Format it so it is visually compelling. In this case I format kinda like when you see inline code in a lot of forums. So invert colors, italicize, that kind of jazz. create 3x as many images in Dall E than I will need. And for the love of FSM, don't make it visually look like every 'hacker image' ever. It is so old, so, I go older! Stained glass, Picasso, art nuveau, wood cuts. Much much better. Put it together and read/experience it at least twice. For areas that don't really hit, go back to the well and have GPT add more flavor. Then do a miniature version of all those steps again. Then again. Output it as three versions: a PDF, and Word document, and a blog post. Triple check the correct license is attached, you spelled your name right, you didn't link to yourself with HTTP as opposed to HTTPS >.< Then publish. This one I started at 8:30pm. I started writing *this post* at 2:15 AM. In that time I took a half hour, 45 minutes to grab some food, have some ice cream, and a cup of coffee. I still need to publish the items, I only have the google doc at this point. But that is a *well understood* task. So here are my metrics: 18 pages at 8.5"x11" Six images Three context summaries All links validated (there are a few I still seem to think could be better) 5 hours, 15 minutes of 'work'. Probably a little less, so let's call it 5 hours. * Efficiency metric: 3.6 pages for hour. This will get incrementally worse in many cases, since three pages are pretty much gimmes. But under 20 pages, I am comfortable with 3.5 pages per hour. * 3727 words. A solid work day when I am in a sustainable groove, I think it is likely I could write 8,000 high quality error free (mostly, let's face it) edited non fiction per day. * I am confident that when get work flow even more optimised of hitting 8,000+ high quality words per day, with images, sources, citations, a 'good enough' literary voice. And, the most important part... not get burnt out. SHAMELESS PLUG Hey, I am unemployed and looking for a job. That's why I started writing again, at first for the sole purpose of documenting the kinda crazy-large corpus of knowledge in my brain pan. My website runs no trackers, no data is ever ever sold, there are things to buy/tips/patreon. Normal stuff really. So I don't feel too bad about plugging stuff, at the end, that can financially tide me over. I mean, I'm gonna do it *anyways* and it is CC Attribution licensed. So don't rip me off, I won't do the same. www.japh.one (main website) www.ecrimeinvestigations.com (a straight up passion project, investigation and dismantling scams that aren't big enough for FBI/SEC/FTC/IRS/Attorneys General, etc) www.completemoderncook.com
Loving the consistency in videos David, you are my #1 source for all things AI!
on god, consistency is exactly what we need with content on this tech, and especially from a god tier expert like David
Hey David, your most average fan here :) When you mentioned a 500% productivity increase that kinda floored me. Mine has been between 500-600% Writing a 513 page long cookbook from nothing, just hands on experience, took me 11*16 hours for the bulk writing, and 3*12 hours for all the little items to get ready for print on demand. That last chunk isn't all that automateable. So that's 212 hours with averages to a metric--npt the actual time, just a metric--to 2 1/4 pages per hour. Including artwork, testing, graphic design, iterative testing, yadda yadda. All of it. Basically 500+ pages in a work month. (I was pulling monumental hours though, that's how it ended up being a little shorter than two weeks) since Friday I've pulled 60 pages, a bunch of original art, my own little frame shop, food blogging, stabilized a bunch of my small business, yard work, and so much more.
I don't think I've gone above, or have any realistic way of going above 600% increase. But even 50% would have been amazing. And a lot of it is because of You and Matt Woof.
AMAZING!!!
As someone who started with a love for writing and reading, and transitioned to coding and DevOps, this video discusses the exact thing I love so much about generative AI: it allows you to play with English in different ways to get better results. I like to think of the way we talk to generative AI as "English as code".
Incredibly rich set of thoughtful and deep ideas. An authentic breath of fresh air. Thanks for sharing. 🙏🏼
This reminds me of when I told GPT, "wrong" in a passive-aggressive manner, then it shut down and said it was done speaking with me. Lol
Bing does that too, it gets triggered so easily and is like "we're done here"
I knew Bing shuts down a convo, but I didn't know ChatGPT does.
Your ideas are ahead of the curve of what is current knowledge and what is discussed.
Cognitive offload and finding the right tone to ask the LLM is something people haven’t really considered in the broad space.
Just a very good collection of coherent ideas.
I have been using Gen A.I. for story writing for a few months now. The Polymorphic code slide would really push things to the limit. I agree with you, but that is alot to get my head around.
With a few hundred words of notes you can get an AI to generate an outline using the Hero's Journey template. A log line, protagonist, antagonist, love interest, some challenges would be enough. Maybe a synopsis.
You may want to do some homework on process improvement techniques.
Hi David, thank you for this one! Did you look at h2oGpt AI? They claim to be an open source organization and want to make fine tuning LLM as simple as possible for the average joe! I think you will find it interesting and if so would be nice if you make a video on them. :)
33:46 any recommendations/links on the fastest path to leveraging tree of thought based context for agendas less boilerplate then a pre-existing and well understood context such as an occupational role? For example, personal multifaceted and standing agendas that are switched between throughout the day. Simplest examples being movement between personal and professional aspirations. Would be great if I could flip to my ____ contextual construct and then field that email that just came in or initiate a new research thread with all the standing context and research to date already in mind and ready to build. A well executed version of this idea could even generate things proactively. Planning to do this and will be building it if it’s not currently able to do what I have in mind. Any immediately viable solutions would be great - if you (Dave) or anyone else out there has any pro tips to pass on. ✌️
Thank you.
Lets goo
Definitely agree
I sent a message on LinkedIn. Great video!
Can you do a video that talks more in-depth about Meta-Programming and Polymorphic App Design?
Still working on the principles, but yes, eventually will do
Thanks David you inspire me a lot!! also videos about LLM linguistics for tech people.
What is the more universal human characteristic fear or laziness?
Will we have websites and apps that need internet access in the future?
I wish i'll see a future where the internet reverts back to raw essential data streama and everyone will have their personal ai to display the data on whatever platform the user is looking at in the perfect way for that person
We'll basically have the Hitchhiker's Guide to the Universe. Fully offline Wikipedia and general AI.
@@DaveShapthe world is too fast changing and transactional for fully offline to be anything more than a backup.
I use the model as a decision maker in my programs. Nothing more. The models are too inconsistent when generating code.
how?
I've had great success with generating code. I think GPT4 (not 3.5) is good enough if you provide context and examples, and can judge the output. I would say 50%+ of the code I commit now was "written" by AI and just orchestrated by me. I really enjoy the workflow because using so much natural language helps me articulate the problem and think through it thoroughly. It's getting both the code and explanation out of my head and onto the screen.
What decisions are you making? It's funny because we seem to be doing the oppisite. I don't trust it to make architectural/important decisions for me etc, just write what I tell it to.
18:54 do you use 32k context btw?
Share Link to slides?
You've been keeping a steady pace of videos about one every day... Careful of burnout, a few ( two or three) a week is good till we pass the knee. Are you looking to hire some help to create content?
thats part of his 5x productivity;-) maybe he can create a pipeline to release them later so he can take time off whenever he wants, while keeping us happy;-)
@@danson3038 So I made this comment before watching the video... So talk about predictive product
agree on almost everything, but I seriously still don't think that we can generate code on demand (29:37). Not as it is. Code must be as deterministic as possible.
9:17 please be kind to your AI 🥹
This comes naturally to older users. Some of it based on caution.
LLM as interlocutor and go fer
David! where is your Star Trek Suit! You are out of uniniform!!!
Don't give away all the secrets haha. Good things coming.
That was amazing David! Well researched and presented and mimics if not expounds and improves upon many of my thoughts and deployments in the space…great job once again and I tip my hat to you sir!🥳🦾🤖👏🏼
damn ... no red shirt today ...
AGI was here 2024
Never been this early
TL;DR Davids etimates remain very much in ballpark to my own experience with writing and generative image creation.
Okay, because I really, really liked this video, I decided to break out the ol' timer, and see how much active creation, effort, formatting, image creation, link checking, hallucination checking, and so on it takes me to create the following:
1. Original detailed technical documentation with ChatGPT4, Plus.
2. All original images with Dall E via bing. Bing lightly watermarks it's images, so those had to be scrubbed.
3. Keep a document of most of the prompts I used. Some I just lightly edited a few times, so my actual number of prompts that I ran were ~3x more than the original prompts.
I developed a couple of documents. I will link them to 'show how the sausage is made'. All GPL, CC, attribution. They allare around Information Security. And of course, if you are uncomfortable clicking a link here I understand. Heh, it is kinda what we infosec folk say NOT to do. (on LinkedIn I'm Jeremy Pickett westcoast, web home is www.japh.one, happy to link my bona fides). This is the document in question:
docs.google.com/document/d/1BIAmeke9vZrl13Iio11xkkc8bnI64lT2/edit?usp=sharing&ouid=104982846581126963187&rtpof=true&sd=true
Subsequent docs will be in replies to myself, so I clutter up main as little as possible.
I grabbed the document/blog post prompt from a running spreadsheet of mine, created by (drum roll) ChatGPT. The prompt was, "Create a detailed specific blog post with four sections for the subject Threat Intelligence and Its Role in Strengthening Blue Teams". That created a response with ~2000 words. That is the seed. So there are actually six sections: Introduction, S I, S. II, SIII, S. IV, and Conclusion.
Each section, while wrangling the responses in a Google doc with a fairly standard template I use a bunch (headers, footers, fonts), I took each section and ran a subsequent prompt. "Can you expand on Section 1, "Section I: Threat Intelligence - The First Line of Cyber Defense" of the document we named INT-1 into details and specific examples or use cases. Include links inline to the details in a natural manner. Use as many tokens, words, and sentences as you can. Explain it like a concise college grad student, do not guess. Link to webpages, urls, sources inline to the blog article."
Then for each section do quite a few things: do the links work? 50% failure rate, so fix those. Is tone, language, fairuse/plagirism okay? Fix those. Is it correct? I've been in information security for 20 years, so I can judge that part really, really fast. Rewrite as needed, and make it sound less like a college freshman writing their first year essay while WAY TOO STONED for the time being. Get through all six.
Then, for each section, try and find a piece of history, subject context, website summary, whatever that can compliment the larger section. Format it so it is visually compelling. In this case I format kinda like when you see inline code in a lot of forums. So invert colors, italicize, that kind of jazz.
create 3x as many images in Dall E than I will need. And for the love of FSM, don't make it visually look like every 'hacker image' ever. It is so old, so, I go older! Stained glass, Picasso, art nuveau, wood cuts. Much much better.
Put it together and read/experience it at least twice. For areas that don't really hit, go back to the well and have GPT add more flavor. Then do a miniature version of all those steps again. Then again.
Output it as three versions: a PDF, and Word document, and a blog post. Triple check the correct license is attached, you spelled your name right, you didn't link to yourself with HTTP as opposed to HTTPS >.< Then publish.
This one I started at 8:30pm. I started writing *this post* at 2:15 AM. In that time I took a half hour, 45 minutes to grab some food, have some ice cream, and a cup of coffee. I still need to publish the items, I only have the google doc at this point. But that is a *well understood* task. So here are my metrics:
18 pages at 8.5"x11"
Six images
Three context summaries
All links validated (there are a few I still seem to think could be better)
5 hours, 15 minutes of 'work'. Probably a little less, so let's call it 5 hours.
* Efficiency metric: 3.6 pages for hour. This will get incrementally worse in many cases, since three pages are pretty much gimmes. But under 20 pages, I am comfortable with 3.5 pages per hour.
* 3727 words. A solid work day when I am in a sustainable groove, I think it is likely I could write 8,000 high quality error free (mostly, let's face it) edited non fiction per day.
* I am confident that when get work flow even more optimised of hitting 8,000+ high quality words per day, with images, sources, citations, a 'good enough' literary voice. And, the most important part... not get burnt out.
SHAMELESS PLUG
Hey, I am unemployed and looking for a job. That's why I started writing again, at first for the sole purpose of documenting the kinda crazy-large corpus of knowledge in my brain pan. My website runs no trackers, no data is ever ever sold, there are things to buy/tips/patreon. Normal stuff really. So I don't feel too bad about plugging stuff, at the end, that can financially tide me over. I mean, I'm gonna do it *anyways* and it is CC Attribution licensed. So don't rip me off, I won't do the same.
www.japh.one (main website)
www.ecrimeinvestigations.com (a straight up passion project, investigation and dismantling scams that aren't big enough for FBI/SEC/FTC/IRS/Attorneys General, etc)
www.completemoderncook.com