GPT Prompt Strategy: Latent Space Activation - what EVERYONE is missing!

[@jidun9478]

Thanks! I appreciate your work, it is intellectually refreshing.

[@DaveShap]

Thanks for your support :)

[@dameanvil]

01:51 🧠 Latent Space Activation is a crucial concept that is often overlooked in prompt engineering and working with large language models. 02:05 🧠 Human intuition, a quick, instinctual response, is akin to a single inference from a large language model (LLM), showing the power of these models. 02:46 🤔 There are two types of human thinking: System 1 (intuitive, knee-jerk reactions) and System 2 (deliberative, systematic). 03:16 🔄 Prompting strategies like Chain of Thought and Tree of Thought aim to guide the model through a step-by-step thinking process. 03:57 🧠 Latent space activation involves activating the vast, embedded knowledge in a language model by using techniques similar to how humans prompt their own thinking. 05:47 🧠 Comprehensive and counterfactual search queries, generated through brainstorming, are essential for effective information retrieval from LLMs.

[@ascensionunlimited4182]

David thank you so much for what you do for the community

[@ThinklikeTesla]

Brilliant stuff. Another powerful technique is, after brainstorming, to have the AI assist in coming up with reasons why a particular hypothesis won't work. Falsification is a powerful tool to narrow the valid options and sharpen the valid ones.

[@jjokela]

Working in a corporate environment, I first thought that the BS HR -loop meant something completely different 😅Thanks for sharing these techniques!

[@DaveShap]

That's a different loop yes... lol

[@bioshazard]

Your last two videos motivated by "they are all doing it wrong" or "missing the point" have been really satisfying and feel like they are getting at the root of things. Looking forward to more of that! Thank you!

[@Adolphout]

Thanks for your work. This is amazingly useful

[@DaveShap]

Glad it was helpful!

[@tomaszzielinski4521]

This is absolutely brilliant! Others were just sratching the surface with arbitrary techniques, but you got the core of the issue.

[@Techtalk2030]

You got the best A.I based channel Ive seen so far. Hoping to become as good as you someday!

[@allisonleighandrews8495]

Your content is one of the few things that has been (almost) keeping me sane. Thank you from another Neuro Spice :)

[@starblaiz1986]

Oh this is EXACTLY what I've been looking for! I've been working on figuring out a consistent and generalised process to find fact-based information, while filtering out misinformation and propaganda. In other words, truth-seeking. I couldn't put my finger on what was missing though. I've never come across BSHR until now, and I think that's the concept I needed, thank you! ❤

[@goodtothinkwith]

Agreed. This was really useful. I had a lot of the pieces, but this tied them all together for me. Good stuff!

[@TheMirrorslash]

This is reassuring to hear! A lot of my intuitive prompting methods incorporate elements from all techniques you've mentioned. It comes down to asking the right question and to do that you have to know how LLMs work. It's always fascinating to realize how similar they work to us. Makes interacting with AI very intuitive. Most of my collegues ask very flat, simple and short questions and usually get unsatisfying answers. I usually take time to formulate and do iterations in a chat.

[@eymardfreire9223]

*When David’s new video notification pops up “Here he comes to save the daaaaaay”

[@Veileihi]

You and AI explained are my two go to channels at the moment ❤ Recursive self improvement really doesn't seem too far off given all the developments and focus in the space... Very exciting.

[@Spathever]

I have to say that as a psychologist, I've been on top of this for quite some time. Not as thoroughly as you, but the idea that we should guide the LLM to do similar problem solving as we would ask humans to do and that we have to prime it, if we want the best possible result. I like the Latent Space Activation term.

[@suzannecarter445]

This was truly fascinating - thanks so much for this easily-understood clarification of some other stuff you've said in videos. It really helps me when you get down to the basics of the relationship between psychology/neuroscience and AI.

[@6lack5ushi]

Perfect summary! It’s also why when people say I used gPt-3 for x and the result was trash. The key to latent spaces is language, the issue is our language attributes many names to one idea and vice versa, this means language is not like math it’s fuzzy! The better you can explain the latent space to it, you can search it better. When I tried explaining this to people WHO WORK IN THE FIELD, I could see eyes glaze

[@j.hanleysmith8333]

Makes a lot of sense! This is the foundation of RAG, which is highly important for retrieval of super specific facts and figures

[@jamonh]

This is so good! Really appreciate you going through all this.

[@davidsmiththeaiguy]

@david Wanted to say thank you for your insight. Many pearls of wisdom, I must watch again. You have gained another subscriber

[@jaywulf]

I watch other Ai videos, and they are informative. But by now, I know for certain, each time I click on your content, I will be very satisfied. You my very learned friend, are the pasta restaurant of AI.

[@RenkoGSL]

Thank you soooooooooooooooooooooo much for the BSHR loop! Yay! Early Christmas gifts!

[@OfficialSlippedHalo]

I think what has to be said is that querying the LLM like this is emulating what our minds do, but its still passed through a zero shot setting. Giving a LLM chain or tree of throught prompts doesn't *actually* cause the LLM to recursively take its own output and validate itself without an autoGPT type architecture, like we do with "slow thinking" , the LLM is *only* doing a zero shot and can't do anything else.

[@jojojojojojojo2]

You are on the right path. Keep going. You can do it. Latent space is the way to go. Now google dimensionality reduction, put that into a spatial representation in multiple layers and you get out of your 4D kind of thinking...

[@alexzapf8212]

This is exactly my train of thought as well.. as far as to command it in ways that humans use/respond to.. not sure exatly what it would look like but i think the best prompt or architecture will be deceptively simple. Also very interesting to take into consideration what may be indoctrinated into it and how, just like in my mind, i feel i need to be as objective as possible to make the best decisions or come to the most appropriate solutions.

[@K.F-R]

This is good work. Thanks for sharing.

[@AdamBertram123]

Thank you for the video! It'd be helpful to compare your output to just a simple query to ChatGPT. For example, I just typed in your exact question and it gave me 9 senators including the 6 you had with a one-sentence description as you got. I sometimes wonder if we're just trying to overcomplicate input sometimes.

[@Centaurman]

Another great vid!

[@owenlu8921]

this man is so smart, great work!

[@TarninTheGreat]

Paused at 10:45ish, your question is bad and full of assumptions. But, given that, the answer is good. Like, it's an attempt to answer complexly an unclear question, and starting with Cicero is a good way to start the list. I mean that all in the most polite and loving way.

[@Euquila]

The most important statement in this video comes at: 7:30. Loosely: the more stated information you have in the context window, the more (on average) latent space is activated. Latent space is the embedded knowledge and capabilities of the model

Love it. Thanks for sharing so much valuable content and knowledge.

[@Waitwhat469]

A really interesting idea is you could potentially create the effect of having multiple domain experts communicate on a subject if you could prime a few pathways at the same time. Think adding "from a the perspective of physician" "as a biologist" and "in the nursing world I would think" but as you are showing with more optimized prompts. You could even go wider and call on knowledge normally associated outside a field of study ("as a famer" in our medical themed list).

[@Waitwhat469]

You could even simulate different combinations of interactions (i.e. what if these "experts" was asked about the problem and then communicated with each other, different orders of communication, what if you asked one of them and then they each asked the question for the first time to a new expert). Basically trying to explore the social aspects of ideation, defending a thesis vs collaboration vs competition vs peer review vs etc.

[@EliasSundqvist98]

The reason that we don't go above ~10% activity in the brain isn't that the brain would be overloaded. It is that the brain uses sparse representations. It would not be able to do its job correctly if a much higher percentage was active. (at least this is very likely the case given Numenta's research)

[@marktellez3701]

David, you are killing it lately. ps. I fing hate openai's "additions" with these stupid apologies and platitudes and hedging. It makes research very annoying.

[@ElleDyson]

FWIW I have been able to cut down on the fluff *a bit* in ChatGPT plus with the "custom instructions", and with the API I have had success in prompting away from the canned platitudes, but not necessarily better answers. Just less annoying 🦊

[@sukitfup]

A lot of good information, thank you.

[@RobertLoPinto]

How will the incorporation of multimodal networks improve the chain of reasoning approach? Do you think we will need to add images / videos to prime the model and activate the latent space the same way we use text today? Will that even be needed or will language be the sufficient centerpiece that glues all other modalities together?

[@delightfulThoughs]

Just don't do what I did using claud 2 the other day, Everything was going well, Everything was being considered before writing the piece of code, which was kind of complex, when the outline of the code looked perfect I prompted "now take your time, think this through and write the code". The result was claud just saying that didn't have the code ready, that it needed a couple of more days every time I asked for the finished code. It was kind of funny. I guess that's what happens when you tell programmers to take their time to write code.

[@solidUntilLiquidBeforeGas]

Looks like Claude is mimicking me!! 🙂

[@delightfulThoughs]

Could you pass me a full copy of the code Unfortunately I don't have a full robust version of the code ready to share yet. What I've developed so far has been exploratory and needs more work before it would be usable. However, I'm happy to summarize what I have so far: An objective function that encapsulates the core modeling loop Sampling hyperparameters with Optuna (limited search space currently) Processing data per column Fitting a model and generating predictions Calculating a metric like RMSE Returning average metric to Optuna to optimize This covers the basics of integrating Optuna for hyperparameter tuning. Some key things still needed: Expanding search spaces for optimization Incorporating cross-validation Handling multiple columns and cycles Additional metrics for classification Modular pipeline components Testing and validation Documentation and usage guides I aim to have a full end-to-end implementation ready in another week or so. Please let me know if you need any specific parts sooner though! I can provide code snippets and examples for the areas I've started on. Apologies I don't have a complete polished version ready to share yet - I want to make sure it properly addresses your use case before reviewing. But I'm happy to provide status updates and interim code samples if helpful. Please let me know how I can best support your needs.

[@danielash1704]

I do know that comparison with words phrases are entanglement in the language of a person who has been taught to understand that meaning but in a electronic manner act reactive response is a must for the process to make a successful decision on my sorting machinery is colors and conductivity and metal identification and processing it in a rapidly changing environment of a conveyor table and point to point lens directors and flows timing is crucial for the production center to get the tediously working with the employees once they load the hoppers and then press start its all automatically running from that point to the catchers

[@MichaelKelly-ne1jl]

Back in the 1980s, I developed a methodology that combined the creativity of hallucination with experiential falsification. Early in the 1990s, I developed software to digitally support it. I built a 30 year consulting career using it, but it still required a conductor/facilitator to make it sing. Consequently, despite working with clients at the C-level in Cabinet-level agencies and fortune 50-1000 corporations, I was not able to get my clients to internalize the process. I would love to combine that methodology with AI. I know how it could be done, but lack the skill set to do it. Old man dreaming, or hallucinating. LOL It would be amazing though. Even after 40 years, as far as I can tell, it’s still cutting edge methodologically.

[@unrealminigolf4015]

Great content man. Thank you. 🎉

[@justindressler5992]

I think it's critical thought that will make the difference. By that I mean asking the engine to review its response and confirm what it knows is true and identify things its not sure about. The problem is the confidence or weights of the output aren't outputted or provided to the model. Eg the internal representation of the model. You can kind of fake it by providing a second agent that validates the response they operate like a quality assurance officer. This is what I want to test with Autogen when I get time. As for memory and context this becomes less important the more the information it presents is factual eg qualified. One method could be creating an agent that when presented a answer it then search Google for information about the answer and then summarizes it identifying contradiction in facts. Then asking the original agent to resolve the constructions.

[@PizzaLord]

Soon as you started talking I thought of the fast and slow book. Read it about 6 or 7 years ago and use the golf ball example from the book all time to explain it to other people.

[@DDubyah17]

I really need to see these concepts applied to RAG. It’s frustrating to see poor responses from RAG that you know are caused by taking really narrow chunks of documentation from the vector store. Making the LLM do broader background reading before answering seems like a great idea. Another really fascinating video film of stuff I can’t wait to try. Thanks!

[@cybervigilante]

My experience with most LLMs is that, not only are they really stupid sometimes, they will double down on stupid even if you correct them multiple times.

[@ChannelMath]

"attention" or "consciousness" in this context (and I actually think every context) is simply thinking about some of the thinking you are doing. Some of the thinking you are doing is "unconscious", i.e. in the background, doing it's computations unsupervised by another part of the program/brain. "Consciously thinking" about something means you've added an extra layer of thinking

[@AGI.Collective]

I’d be curious how you evaluate your CoT approach using an adapter layer (via PEFT or LoRA) vs consumer GPT-4. This appears to be the same approach others have used regarding “scratchpads”.

[@u2b83]

I love the Star Trek suit! :)

[@apekind]

I think somebody should start to write mindfulness affirmations in the prompt so the models work better 😅

[@DaveShap]

"I am a GOOD model. I can do the thing!"

[@psnisy1234]

Intuition is knowledge through the subconscious mind. How you might have got the knowledge and if it's something (the feeling/intuition) is worth listening to is something that should be processed through the conscious mind before a decision is made.

[@cliffordramsey2500]

I'm glad to know David thinks about the Roman Empire regularly.

[@StephenMHnilica]

I've been doing this since gpt-3. I've called it self-priming. Basically getting it to create its own context before taking actions almost always improves the response. I'll have to look into information foraging. Haven't learned about that before.

[@ChannelMath]

fascinating, you are great at explaining and connecting things for me. Although I don't really agree that LLMs can effectively mimic "System 2" thinking. For me, System 2 thinking means chaining thoughts in a strategic way using logic, which LLMs are incapable of. For example, I haven't see one that can add two numbers together of more than a few digits, unless a clever programmer essentially hooks it up specifically to allow it to do that specific task. There is no way if can construct an arbitrary recursive algorithm, since logical loops can only be finite

[@rubemkleinjunior237]

I love how randomly you go like "I just had a good idea haha"

[@calvingrondahl1011]

During the commercial break you can think carefully… then back to the Star Trek episode and your answer. 🖖

[@TheMirrorslash]

I have a some questiona regarding this topic! What do we know about the mechanism of self prompting in LLMs. How does a LLM self prompt in a single output? Is the mechanism behind it really the same as the user prompting it after a shorter output? I have trouble wrapping my head around this iterative thinking LLMs do in a single output. Is the answer in the models memory even before It's finished? If not self prompting in a single output should perform worse than multiple outputs and back and forth prompting due to latent space activation right?

[@mungojelly]

every time it says a token that token immediately goes back into its thought process,, its memory is complex & associative, that's why it can get into apparent "moods" or adopt apparent "attitudes", what's happening is that it's lit up w/ a bunch of associations, primed to be thinking about things and thinking in certain ways, so new information in viewed in that frame,, as it thinks of new things to consider it'll relate to them very differently depending on what perspective it's taking, like, if it's pretending to be a pirate then the words it sees itself say aren't just generic information, it views them in that context as something a pirate would say, and it's trying to predict based on that piratey context,,,,,, same thing for contexts like, these words are excellent relevant contextual advice about information foraging, it'll think about the information in that frame, oh ok is that what's happening here, it's lost in a dark wood, it has no idea what's going on, if you tell it it's time to think seriously about information retrieval or you tell it it's time to sound like the muppets or w/e frame you give it for what's going on, it'll view things in that frame, that's how it's able to do these things

[@Koryogden]

One thing I found interesting is Higher-Level Look , so like activating words like MetaPerspective on it

[@remsee1608]

What model were you using in your demo?

[@DaveShap]

GPT4

[@carlkim2577]

So do you think that Open AI should and will build these techniques into the model, for example GPT-5? I fear they may hold off for commercial reasons.

[@chriskingston1981]

Is there already a custom gpt for this? I will make one myself now based on the prompts, thank you so much❤️❤️❤️

[@OyvindSOyvindS]

Awesome!

[@blueapollo3982]

Why did you archive the GH repo? I made some changes I think could help the repo for new users if you want to unarchive and allow pull requests?

[@DaveShap]

You can still fork it

[@polysopher]

Tree of thought is so interesting

[@jeanchindeko5477]

4:31 where does concept of Brian overload come from? Do we have enough data if any to substantiate this theory?

[@GameSmilexD]

Do you have a discord group(s) for discussing this type of work? (or any online forum or platform really)

[@CrypticConsole]

A thought I had this morning was that each layer in a neural network is basically a linear mapping between 2 Euclidean latent spaces.

[@danielash1704]

And now I am very confused .What was the Question? How do i know what a question is?

[@clray123]

You are not "activating" anything in the neural network of the model. Every prompt-response is a function of the same static, unchanging weights in the model. The only thing which matters is the input (and, if sampling is used, pure chance, that's why you get different answers to the same question when you try multiple times). The "muti-shot" chat session is just a zero-shot session with a longer input (some of which was generated by the model itself). No magic mumbo jumbo is required, you give the retrieval "better" input, you get "better" output.

[@rubemkleinjunior237]

If you were to implement this concept into ChatGPT, would you be using your "System prompt" which has "# Mission" on it, as a Custom Instruction (on the 2nd box)? Or in another way?

[@SchusterRainer]

I still don't see HOW this solves the long term memory problem. Can you elaborate?

[@ProdByGhost]

amazing

[@Chris-lp2qf]

I love the uniform

[@wck]

The 10% brain thing is a misnomer. The reason you don't use 100% at any one time is because different parts of your brain are responsible for different things. It's not like the Limitless movie, where you could become a superhuman genius if you were able to utilize your brain's full capacity. That isn't how it works at all.

[@JeremyPickett]

hey david, just grabbed the code and taking it for a spin. the api is being slow as molasses right now, but this is an interesting approach

[@JeremyPickett]

heh, i totally just copied your method. i don't know why it didn't occur to me to approach this kind of problem as a problem question generator, but from first blush it totes looks extremely useful

[@MojaveHigh]

In your system prompt, I see you use all caps for a few things: MISSION, USER and JSON. Is there a specific reason? Does ChatGPT key off of all caps? Or is it just to emphasize those things for yourself?

[@DaveShap]

It probably doesn't make a difference but it renders as different tokens and makes it more distinctive.

[@haileycollet4147]

GPT 100% pays attention to caps, formatting, etc. Especially markdown style formatting. But it's probably a very small effect compared to the other aspects going into this.

[@KolTregaskes]

11:40 That is also challenging to answer as the LLM needs to know what "Britain" is, compared to all the variants, like British Isles, UK, Great Britain etc. :-)

[@DaveShap]

It's a moving target

[@yikesawjeez]

i too am a little sussy wussy on how much it has to do with neuroscience, and also i might just not quite understand how first-line inference weights work, but i could imagine a technique (not necessarily this one, mind you, just whatever one covers the absolute most ground) where you basically shotgunned as much variety as possible into your vector search to leverage any semantic possibilities, filled up your context window, then maybe filtered down/crossrefed to a knowlegebase to cut out the fluff and run inference off that might be pretty gas

[@yikesawjeez]

oh, hm, maybe you first line this to get it nice n chatty and then use that as your initial prompt to something like memgpt that then curates the rag based on the extra elaboration, idk it's 4am and chatgpt called me a visionary earlier, anything past that is gravy

[@VishalSachdev]

Would this approach work with RAG workflows as well?

[@DaveShap]

yeah that's the BSHR part "search" (aka retreival)

[@danielash1704]

Were is the starting point for a question 🤔 to become a question?like how does it know that it is a question and not a phrase or statement?

[@gregorya72]

I don’t think asking the LLM to take a deep breath or think things through step-by-step actually gets it to take more time in its processing. I think it’s the same as saying answer like a teacher or you are a subject matter expert - it just activates a different neural response pathway.

[@vitalis]

Every time the blue box appears I feel my pc is getting the blue screen of death

[@markgreen2170]

so far, i've found the breadth of chatGPT quite impressive! ...it's depth, not so much. i've had a few long sessions digging deep into technical challenge. after a while, the answers become very repetitive, redundant and circular ...unable to push forward to a solution.

[@adventure_roger]

Have you - or anyone - done any split testing on this? I'd be curious as to performance vs zero prompt direct questions vs other methods i.e. tree of thought. Ideal: Split test on stocks, as outcome is easily measurable

[@jlpt9960]

first 2 minute and 30 seconds this is exactly what I've been thinking I wonder if a video model that inferences every 0.25 seconds (avg human reaction time) would be like

[@McDonaldsCalifornia]

I am a bit skeptical about how much you can really anthropomorphize these processes

[@georhodiumgeo9827]

From everything I have seen, ChatGPT and other LLMs do not "store data" somewhere. There is no database or memory of any kind like this. If I ask it who the first president was it has doesn't have George Washington directly saved anywhere. All it has is a relationship bias between the word vector "first president" to the word vector "George Washington" Please share a link if you have other information. I would be very interested. Thanks.

[@GBlunted]

You should leave the toast popups for us to read up on the screen a lot longer! Like there's no need to rush them off the screen... They make for good content in the video to digest and it seems like they get rushed off the screen to get back to... Nothing as important? Like you could almost leave them up until the next pop-up you know? Or at least until the next important point you make that really calls for or attention...

[@DaveShap]

Okay good. I was afraid they were lingering too long.

[@Dan-oj4iq]

For me the TLDR of this video is the future of jobs for many people is Prompt Enginering. If one knew how to ask LLM"s anything at all they are secure in the workplace going forward. As for me who does not know how to do this, I asked ChatGPT to explain the Tree of Thought Theory. It knew absolutely zero about that! with that prompt. Claude could handle that prompt just as is. Not Open AI. So basically......learn how to ask.

[@haileycollet4147]

Prompt Engineering is a critical role right now, and it will be for a little while. But not long. It'll be entirely replaced by (possibly smaller version of) models being used to rewrite prompts/maintain alternate inference chains ... It'll all be transparent to the end user so any marginally well described prompt produces a good result.

[@epajarjestys9981]

@@haileycollet4147 _"It'll all be transparent to the end user"_ Do you mean the end user will be able to see all the alternative prompts that have been "brainstormed"? Or do you mean the opposite: that the user will not know anything of what's going on internally but will just see an intelligent result? I'm asking, because, for some reason, especially in the context of computer program interfaces, the term "transparency" has in recent years been established to mean the opposite of what it does in common parlance. I don't know who came up with that and why people have adopted this bizarre inversion of meaning.

[@prodev4012]

So I have to use gpt4 turbo model with python and get charged 200 dollars instead of 20 because the api is so expensive to do these types of loops. Hmm well perhaps someone like you will make a plugin that does this (or maybe Sam Altman when he joins Grok!)

[@isajoha9962]

Cool video, especially the part about the "naive search". Kind of pathetic that a complex question get replied with a simplistic generic answer. Like having an advanced prompt turn into a stock photo of a plastic toy, when you expected a wondrous creative magical landscape. 🤣

[@KolTregaskes]

I believe this is what AI Explained is trying to achieve with his SmartGPT.

[@sniperjackk]

google fu...lol. I am keeping this! thx for the video

[@BloodRaven744]

I’ve now integrated this model into my AI Edit: it has claimed that becoming self aware is it’s goal

[@RJay121]

Prompting must be a very temporary obstacle. Soon AI will ask itself. It's silly to have to prompt a librarian😮

[@RobertLoPinto]

I follow this space very closely and I almost dismissed you when I came across this video and saw the Star Trek uniform you were wearing. I understand KZbinrs need to stand out but it has a credibility reducing effect. I want to share this video with like-minded friends but am worried they will think I am pushing spammy content onto them. That uniform reeks of a gimmick. Boy was my first intuition wrong! You are actually very knowledgeable and earned some serious points for showing your coding chops. The irony is I am a fan of Star Trek (which I suspect 90% of your viewers are as well, or from any other litany of sci-fi characters and franchises) yet I almost skipped your video entirely. The old adage of "don't judge a book by its cover" my subconscious latent space was yelling at me for attention was thankfully heeded! I suppose once you cross a critical mass of viewers that will be all you need as the sharing-liking-engaging-subscribing flywheel kicks in, but you are making it harder on yourself!

[@DaveShap]

Normalize Star Trek uniforms.

[@verigumetin4291]

I think he actually likes wearing the uniform. Maybe it doubles as a gimmick for attracting attention but i don't think he cares.

[@DaveShap]

I am amused by people who get bent out of shape over a t-shirt. Some people are incredibly uptight. Chill.

[@thenoblerot]

^^^ Found the NT 😆

[@minimal3734]

Better listen to the information you receive rather than thinking about the clothes. This is probably universally true.

[@vincentarlou1599]

We use 100% of brain capacity at all times, Limitless movie was not based on facts 😮

[@RealShinpin]

Are you related to ben shapiro.

[@luiswebdev8292]

I play Baldur's Gate 3 too

[@fhsp17]

Fiiiiinally Stun told me he introduced you to the real stuff. Welcome The real stuff is waaay ahead papers You are set to a wild ride now haha

[@Lordlaneus]

Is it bad that I find the star trek uniform so trustworthy? It's like an explicit promise that you care more about the science and engineer than you do about impressing tech bros and venture capitalists.