Unlocking Online Signature Legitimateness for Businesses in the European Union

How to eSign a document: online signature legitimateness for businesses in European Union

I am a Golden Gate Bridge is a meme that sprouted from the latest LM interpretability research published by anthropic while it may not be a popular meme the details the researchers discovered by having Claude saw it output anything and everything of itself being the Golden Gate Bridge is a pretty huge step towards reading the mind of AI and no they did not prompt the AI to say so instead they achieved this by tinkering the Neons that the AI model has or by using yet another AI model to introduce themselves as the Golden Gate Bridge wait a second using AI to explain AI isn't that a bit counterintuitive well actually there are neural networks that are easier to interpret than the others but before that why are AI models hard to interpret in the first place for one reason you might not be looking at them at the right angle so if you do want to completely realize the true capability of chat gbt and not just a fraction of its power HubSpot has you covered this time HubSpot is offering free resources for using chat gbt which is is a comprehensive guide to enhance your work productivity with AI this is a perfect guide for anyone that wants to get into improving their working efficiency while keeping up on these rapidly growing AI Technologies aside from the 37 page in-depth guide that they provide for free there are also other useful visualizations such as chat gbt flowcharts instruction templates content refinement checklist and a few more my favorite part is the 100 prompts in here prompting is so important that getting a head start on it will help you a lot so yeah I highly recommend you to check out the chat gbt bundle using the link down in the description to level up on your chat gbt skills and thank you HubSpot for sponsoring this video anyways you see the whole reason why neural networks are so popular is because they can approximate any function without actually knowing them or by finding the patterns from the data for example if you train a dog or cat classifier what is really happening under the hood is that the network is iteratively trying to adjust its own parameters by comparing the ground truth and prediction to learn whether the input is a cat or a dog which would create a really complex underlying function in the process so the size of the network determines how complex the function can become aka the expressive power of the network from a higher level perspective each neuron may represent a feature of the image that will help to determine if it's a cat or not but as the task gets more complex a neuron in the neural network would have to multitask and fire on different occasions for example instead of having a feature for one neuron there's a combo of neurons firing th determine one feature and this is what they call the superposition hypothesis which is the ability of a neural network layer with Dimension n to linearly represent more than n features so when a network is as big as having a few billion parameters I think you can kind of see now how it is hard to interpret the firing of each neuron as it can probably mean a billion other things when it fires with the possibility of being involved with the trillion of other combos this is why researchers call these neurons polysemantic meaning that they respond to multiple features and are often really unrelated like cat ears or Cloud so the goal the researchers over at anthropic is to make these features monos semantic which only responds to exactly one feature and with that we can then kind of read ai's mind around eight months ago they published research where they saw the potential in their method which worked on their toy language model that is composed of one layer Transformer the method they propose is to use an arch leure code a sparse autoencoder to decompose the activations of a model into more interpretable pieces by training them on the activation patterns so the whole concept of an auto encoder is that there's an encoder that learns how to compress the data then there's a decoder that learns how to decompress the data this data reconstruction process lets the auto encoder learn the activation patterns when a given input is used but that would still make the neurons polysemantic as every neuron in the outter encoder can still have a say about the prediction so this is where the sparse constraint comes in by adding this term as a penalty in the loss function it forces the network to only activate a handful of neurons while learning and ignore the others which creates sparsity this makes the activation of SAE a lot more obvious than can be used to interpret the activation patterns of large language models they call this dictionary learning where one activation would refer to exactly one feature like how you can search up a meaning in a dictionary so using the same technique this time they have tested on an actual production ready model to see if it scales while we don't know how big son it is which is a key information for interpretability research that they decided to leave out due to Commercial and safety reasons we do know that they trained three different sizes of SAE and for all three of these saes the average number of features active on a given was less than 300 and the largest SAE was able to explain a to 65% of the model activations from Claude Sonet which is really impressive the results they got are really cool too firstly they tested on a few features like when you mention Golden Gate Bridge the greatest activations are essentially all the text directly mentioning the bridge and weaker activations also include texts related to tourist attractions similar Bridges and other monuments in San Francisco while this might be a flaw where the SAE is not able to extract and discriminate among features as cleanly it does show an interesting aspect where the feature is shown to have different amounts of connections with other words then to confirm if the feature actually means something they would dial up a feature 10 times where if you then ask Claw on it what is your physical form it will always reply it is the Golden Gate Bridge so good news this actually works but what's even cooler is that they found features that are abstract Concepts and not just some relatively simple abstract ideas like tourist attractions they are talking about features that demonstrate depth and Clarity of understanding which smaller SAE models cannot extrapolate there's this one feature they found corresponds to whenever there's a bug in a piece of code and in hindsight it could just be a feature that detects in taxs and misspelling or a specific python feature but it is certainly suspicious so they dug in more then they discovered that this feature also fires on similar bugs in other programming languages like C and scheme so to check whether or not this is more like a typo feature they tested on examples of tyos in English pros and found that it does not fire in those so is this feature a typo detector for codes well not really because they also found out that it fires on erroneous expressions like divide by zero invalid input and function calls array overflow asserting provably false claims and many more so to double check it they then dialed this feature up and the model proceed to generate an error response where the input code is bug free and by suppressing this feature the model would take the code with a bug and produce a bug-free version of it so they conclude that it is a feature specifically related to code errors and it just shows how deep of connection the model has with some complex Concepts on the other hand they were able to measure these features and create a nearest neighbor map with other features what is interesting about this is that you can see the variations different SAE sizes have when spitting the features for instance a San Francisco feature in 1 million SAE is split into two features in the 4 million SAE and for 34 million SAE it was split into 11 fine grain features the distance between the feature points is also roughly their distance in the concept space and at greater distances we also see features related in more abstract ways like features corresponding to tourist attractions in other regions so with the same logic they are able to identify features that can be used for AI safety they found features like unsafe code bias copany deception power seeking and dangerous or criminal information that only activates on these topics so by influencing them they can dial the feature to make the model as friendly as it can or as hateful or racist as possible which is pretty crazy they didn't show the examples of course but they did mention that it also exhibited signs of self- loath when it was dialed up 20 times they said this phenomenon was unnerving due to the offensive content and the model self-criticism suggesting an internal conflict of sorts but I think there might just be connections between the concept of internal conflict and hate speech and whoever wrote this might have overthink it as a few sections ago in the same blog they were just talking about the internal conflict feature so anyway what does this all mean for AI safety since this is the whole reason why anthropic is doing interpretability research well this does hint at the possibility where the guardrail in the future may not be reinforced with prompting and features can be optimally suppressed instead which would make jailbreak kind of impossible but that is if they are able to scale an SAE model big enough to interpret the state-of-the-art model for us and to be honest that kind of wouldn't make sense financially as it'll be too expensive and would take too long to train while only be able to stay fixed to a version of an llm model further fine-tuning of the L may screw up the SAE so it may be potentially impractical unless the SAE is used to train something else like a reward model which might make sense on the other hand they did not find a feature for hallucination either which is a key problem that LM has so maybe hallucinations are native functions for LMS also there are up to 65% of dead features which they cannot capture with their largest SAE which means the entire thing is far from being completely interpretable it might have to do with cross layer superpositions where a feature might have been smeared across Transformer layers which is beyond what the SAE they proposed could cover and even anthropic kind of indirectly admitted they have no idea how this mechanistic research they've done can be fitted into practical use for AI safety so it really ain't on you if you feel confused after reading this research it definitely feels like a paper where they're just cherry-picking and eyeballing results but I guess you need to design a probe first before you can devise a solution on the other hand if we stem from the platonic representation hypothesis I discussed in my last video maybe fixing data will heal these safety issues and hallucinations themselves and make the model more steerable this could probably provide the most financially efficient outcome than creating these SAE models too however we still got to give it to them for sharing this research because even if interpretability turns out not to be fruitful in the longan run for AI safety this proves that AI is not just something made out of fluke and does actually understand complex logic hopefully in the next update from them we will see some news on the potential applications of mechanistic interpretability on AI safety with their saes because while it may be a bit greedy to tackle AI safety entirely from it having another tool in the shed is always a great idea I skipped out a lot of big technical details about the research like how they use Claude Opus to measure whether a concept is present on a feature activation or not so you should definitely check out their full paper yourself if you're interested thank you guys for watching if you like the video today you should also check out my newsletter where we break down the top research papers weekly the goal of making it easy to read fast to digest and straight to the point as I probably won't have enough time to cover all the incredible papers that are coming out with my videos this would be a great alternative for you if you want to stay up to date with the latest research breakthroughs a big shout out to Andre chelas Chris Leo Alex J Dean Alex marce mulim fifel Robert zasa and many others that support me through patreon or YouTube but follow my Twitter if you haven't and I'll see yall in the next one