正在加载视频...

视频加载失败

Around 3 weeks ago introduced Laya, the first open-source version of JEV, and it has reached a very large scale. Today I am open-sourcing a new physics-based typed decision model called Vega. The interesting parts. It is only 800 million parameters (4B is also there), with 73k token support,...

85,479 次观看 • 1 个月前 •via X (Twitter)

36 条评论

Mohit Kale 的头像
Mohit Kale1 个月前

Incredible to see Indian soloprenuer doing so many stuffs with no huge backing. It's an inspiration. Maybe the next model could be around motion graphics and some sort of diffusion model for media?

Sam 的头像
Sam1 个月前

open sourcing two heavy projects back to back is wild. how are u handling the infra for community feedback?

ata 的头像
ata1 个月前

really interesting! i’m curious whether the physics-based dynamics actually provide an advantage over a conventional classification head. have you tested both using the same backbone and training data? i’d love to know whether the ball dynamics improve decision quality or calibration, beyond just being an alternative way to compute the scores.

Nandakishor m 的头像
Nandakishor m1 个月前

Yes. Benchmarked with multiple datasets.

ata 的头像
ata1 个月前

nice! but have you specifically run an ablation replacing the physics dynamics with a standard classification head, while keeping the same backbone and training data? i’m curious how much improvement comes from the physics itself rather than the pretrained model.

Nandakishor m 的头像
Nandakishor m1 个月前

Ablation studies already done. Will push it in arxiv. And damn the arxiv , they ain't publishing anything..

Harikesh Singh 的头像
Harikesh Singh1 个月前

is there any outcome based difference in Laya and vega or just there is difference in underhood decision making technology

Nandakishor m 的头像
Nandakishor m1 个月前

The architecture is new, training new and output is like most decision model. It's a whole new generation of model

Harikesh Singh 的头像
Harikesh Singh1 个月前

Thanks ...I will try it🙂.

Aravind 的头像
Aravind1 个月前

Btw it’s great to see bangers one after another from you! I worked forking laya to layax! Essentially added 1.Late-interaction option heads: options go through a cached tower outside head_max_len, scored by maxsim against state tokens. 2.A competence head predicting answer correctness, with split-conformal abstention.

Vinay Parisa 的头像
Vinay Parisa1 个月前

This is brilliant and thank you

Kneeanderthul 的头像
Kneeanderthul1 个月前

THE MAN KEEPS BRINGING THE HITS! W

TheBTCaveMan. 的头像
TheBTCaveMan.1 个月前

@saushank_ @openmayhem_ai you already know!!

The Klingon Yellow 的头像
The Klingon Yellow1 个月前

I have an agent trying this out now and building some expertise before adding it to my /system-one skill in Laya is already there.

Harold Finch 的头像
Harold Finch1 个月前

Trying this !

Vijay Rathi विजय राठी 的头像
Vijay Rathi विजय राठी1 个月前

Congratulations and best of luck ! keep innovating and keep inspiring

Aravind 的头像
Aravind1 个月前

Sounds awesome playing pinball with probability to get answers !

bb63 的头像
bb631 个月前

When to use this over laya/jev/openai decision model?

Nasruddin Shah 的头像
Nasruddin Shah1 个月前

adipoli

Kneeanderthul 的头像
Kneeanderthul1 个月前

Where the is @huggingface retweet !? Here's a whole new architecture Might be worth showcasing something so beastly from the creator of the OG decision model.

Rahul bansal 👀 的头像
Rahul bansal 👀1 个月前

He is at frontier of frontier.

Arvind 的头像
Arvind1 个月前

Amazing idea and execution man!! 💪💪

P 的头像
P1 个月前

This should make waves!!

idk 的头像
idk1 个月前

All the best sir. Please keep doing it. We really need people like you.

Lovi 的头像
Lovi1 个月前

Carry on bro. Proud of u

Ajeet Kumar Upadhyay 的头像
Ajeet Kumar Upadhyay1 个月前

Physics-engine inference is not something you see everyday. Incredible work open-sourcing this! 🔥

Georgy GLuhov 的头像
Georgy GLuhov1 个月前

I’ll try it in my RimWorld colony that built with Laya now! Image support it’s perfect for this task, big update soon…

MD Fazal Mustafa 的头像
MD Fazal Mustafa1 个月前

Sick 🔥🔥🔥

Gard Labs 的头像
Gard Labs1 个月前

Create your community bruh when you are still there.

Vibhor Sharma 的头像
Vibhor Sharma1 个月前

my goat is back

NEHAL HASNAIN 的头像
NEHAL HASNAIN1 个月前

Great. Thank you.

Shravan Pendem (SP) 的头像
Shravan Pendem (SP)1 个月前

Can you expand on the training part? What data have you used to train it?

RAVINDRA CHOWDARY ✨🧎🥵 的头像
RAVINDRA CHOWDARY ✨🧎🥵1 个月前

GREAT WORKK MAN!!

Priyam 的头像
Priyam1 个月前

Compared to physical based world models coming into the play where there are so many papers atleast on predictions over physics-based systems. Where do u think decision models/classification models stand here

Samantha 的头像
Samantha1 个月前

Hey Nandakishor, we’re putting together our annual AI & technology feature and would love your perspective on how AI is shaping the future of software and technology. If interested, feel free to reach out.

Nikhil Sharma 的头像
Nikhil Sharma1 个月前

It will be a great tool for verifying prompt injections.

相关视频

Jev is cool. So is it's OSS companion, Laya. The Latest Cool Thing In AI™ tends to get a lot of hype, sometimes without everyone even understanding it. So... what is this thing? Jev is an AI model that consumes input and produces output VERY differently than chat, claude, grok. The input is two things: 1) Text state to assess. Email, html, code, whatever. 2) A set of questions which will be asked about the attached state. The canonical example from TypeSafe's docs is to identify the urgency of a support ticket. We pass the model the customer text + a single noul question "is this urgent?". Jev returns a full set of JSON. This JSON is not generated with token-by-token autoregression. Jev is not trained to produce sequences of text tokens, rather to answer questions, and guarantees well-formed responses. In the example below, we see it produces a 0.99 probability (on a 0-1.0 scale) that the answer is "yes." Jev supports exactly three types of questions (seconds example in video): a) Noul: 0–1 probability that the answer to a yes/no question is "yes." b) Choice: Ask question with pre-defined set of answers. Jev chooses the best and assigns probabilities to each. c) Score: Ask question with pre-defined scale of answers. Jev produces a position on the scale. Jev computes answers for all questions in parallel, making responses super fast even for many questions in a single request. This might seem like a narrow set of capabilities, but in the right contexts leads to incredible potential. It also makes for a useful API / primitive for programming, since the outputs are... *ahem*... type-safe and predictable in structure. Jev is not going to replace LLMs for writing your code, auto-generating your docs, or being at the core of an agent harness. But Jev IS incredibly cool, and will be used to build a lot of amazing tech. Hope this helps.

Ben Dicken

42,183 次观看 • 1 个月前

Pressing, transitions and goals in Canada! After the hiring of Jesse Marsch this week by Canada Soccer , there has been lots of talk about pressing and transitions in the football community. This is THE style of play that we have used for the last 5 years with Alliance United FC in @L1OMens and will give you some insight in how this can look. We went away from the 'possession' based 1-4-3-3 that everyone has been using as the 'best way' to develop players and win games. Many say this is the 'right way' of playing. However, we use the players' abilities and inabilities as the starting point and based on players produced in Canada, within our youth system, this is the way we think will get us the best results and help our players get to the next level. You need to understand youth development in our country to understand what is and what is not being taught to young players. This is not a negative outlook, it is just looking at the reality. Many mistake this style of play as just non-stop pressing and chaos. Yes, these two aspects are important but it is a much more deliberate and planned way of playing that is used to predict where the opposition will play the ball, where the ball can be dictated and what areas the opposition leave open for transitions when they attack. In 2019, I was introduced to Ernst Tanner former Academy Director of Red Bull Salzburg and current Philadelphia Union Sporting Director that has led the club to being one of the top clubs in MLS and arguably the best academies in North America the last few years. Both organisations are known for their pressing and fast transitions in addition to producing top players. He became a mentor to me and gave me insights and education on this style of play. He has changed the way I look at football. Every season we try to recruit players that can play this style that I will describe below and we do not waiver in the way we played. Every game since 2019 Alliance United FC have played either in a 1-4-2-2-2 or a 1-4-4-2 midfield diamond....every game. We have no secrets in how we play. We focus on defending, pressing, forcing teams into mistakes and transition football. It is important to understand that we do not care how much possession we have and we actually want the opposition to have the ball most of the time. Stats in football show that only when a team has the ball 70% or more in a match their is a correlation to winning. Anything less is not correlated to winning in a certain match. In addition, 80% of goals are scored under 5 passes or less and under 10 seconds when regaining possession. We follow the trend! So, here are some important points based on the video: a) Pressing higher up the pitch when possible. All 11 players are committed to the team intention (principles) and if anybody is not committed and does not contribute to this style of play, they do not play. There is no leeway on this. Either you are in 100% in or you are not. All it takes is one player to not commit and the plan will not work. If we do not win the ball on the high press, everyone is to drop behind the ball as fast as possible to restart the press closer to our goal. b) Based on the opponent's scouting we press certain players and decide if will press closer to the sideline and 'pin' the player to the sideline or we dictate the passes and dribbles centrally and 'surround' the player to win the ball in central positions. We also decide if we will sometimes drop lower for the opposition to advance so we can play behind them when winning the ball. This is done if we scout that the opposition centre-backs are slower than our two strikers. This means we will 'outrun' them in the space they leave behind. c) When winning the ball we want to exploit space behind their backline as the priority which means that a player(s) need to be passing options behind the backline and preferably centrally which is closer to the goal, players winning the ball has to look to play the ball to the player furthest up the pitch. Possession is not a priority but scoring goals as fast as possible is. d) When winning the ball we stay as central as possible with passes and dribbling. We use the width of the penalty area (44 yards) as the preferable dimension when transitioning. We want to stay within this width. The wider the team plays the ball, the more time the opposition has to block the middle. PLAY AS VERTICALLY AS POSSIBLE! The emphasis is to get the ball to the two strikers and play 1v1 against the CBs. e) We know that most youth players grow up playing against a 1-4-3-3 which means there is one central striker meaning the centre-backs one back press and one can cover. Against 2 strikers this is not possible and many centre-backs do not have the ability to play 1v1 with space behind them. They are not taught the cues to step and press or drop to protect space behind them. They are never taught at youth level. We exploit this deficiency. e) All 11 players must get up the pitch to close the spaces if we lose the ball from our attacking transition so the opposition cannot transition against us and we can counter-press. f) A goal is the best outcome based on the objective of attacking but at the minimum we want to get a shot on goal from a transition. This is a quick overview of a different way of playing football that Canada will see with the Men's National Team.

Ilya Orlov

14,134 次观看 • 2 年前

How can you solve complex tasks using a Large Language Model? Here is a 2-minute introduction to everything you need to know to 10x the quality of your results. Let's talk about three techniques, in order of complexity, starting with the easiest one: • In-Context Learning • Indexing + In-Context Learning • Fine-tuning In-Context Learning The team that trained GPT-3 found something they couldn't explain: You can condition a model using examples of how you want it to behave. I included an example prompt in the attached video. You can "teach" the model how you want it to interpret questions, select the correct answers, and format the results by giving a few examples. You can also give specific knowledge to the model that will be helpful when formulating answers. We call this approach "grounding the model." There's another example in the video. Indexing + In-Context Learning Unfortunately, there is a limit to how much data you can include in a prompt. We call this the "context size." One version of GPT-4 supports a context of approximately 6,000 words, while the other supports 25,000 words. Although this sounds like a lot, many applications need more than that. Imagine you wrote a book and want to build an application to answer any questions about your story. What happens if your book is longer than the context? That's where Indexing comes in. Using a model, you can turn every book passage into an embedding. These are vectors, numbers that "encode" the passage's text. You can then store these embeddings in a particular database that supports fast retrieval of these vectors. You can then turn any question into an embedding and search the database for the list of passages that are similar to that query. Instead of using the entire book to ask the model, you can now use the relevant passages as in-context information, effectively working around the context size limitation. Fine-tuning Fine-tuning can give you an extra boost to get reliable outputs from your LLM. It is, however, the most complex approach on the list. There are different approaches to fine-tuning a model with your data. A popular technique is to process your data with your LLM and use the outputs to train a new classifier that solves your specific task. Notice that here you aren't modifying the LLM. Instead, you are chaining it with your trained classifier. Another approach is to modify the parameters of the LLM using your data. Think of this as "rewiring" the model in a way that solves your particular task. The results and costs will vary depending on how many layers you want to fine-tune from the original model. Many companies think that fine-tuning is the solution to their problems. In my experience, many will benefit from exploring the other two approaches. I love explaining Machine Learning and Artificial Intelligence ideas. If you enjoy in-depth content like this, follow me Santiago so you don't miss what comes next.

Santiago

384,573 次观看 • 3 年前

David Friedberg: Frontier Models are Training on Your Novel Insights as “De-Identified Data” @jason: “Should they trust any of these LLMs with their proprietary knowledge for fear of having it cribbed into a core LLM?” david friedberg: “I have had experiences where we've asked some fairly novel scientific questions, and (the AI model) identifies it as a novel insight. It's like, ‘Oh, never thought about that, interesting, blah, blah, blah.’ And then using a different account, asking the next version (of the model) later, I've now experienced this. It's like, ‘Oh, well, you could do this,’ and it actually just describes this exact thing that we had in our chat in the previous version. Now, these are a handful of anecdotal experiences, but I know the domain that we work in, and the niche of it, and the ideation of this stuff, and the novelty of this stuff, and the lack of papers being published, and so on. So I know that there isn't some new corpus of information out there that's training the new model. So all I can say at that point is that my conversation or our analyses have been used for training.” David Sacks: “Okay, this does raise a really good question. What does it mean that the model is allowed to train on unidentifiable data?” Friedberg: “Well, that's my point. So it doesn't use any of my personal information, but it can use an insight derived from our chat, which it can then say is some training data that is unrelated. But the truth is, it's actually a piece of IP that's our organization’s IP, and our engagement back and forth. We don't have any NDA or confidentiality provisions or protections with them being a service provider back to us. This is why I care a lot about open source because I don't want them having my chat logs because they can use it for training to create an IP advantage that is now diffused to the rest of the market.”

The All-In Podcast

56,782 次观看 • 2 个月前

Chamath: Two terms you need to pay attention to in AI are Prefill and Decode “There's two terms that I think you're going to hear a ton about over these next few years.” “The first term is prefill, and the next is decode.” “What prefill and decode are, are two very distinct ways of how models think, and how a model goes through the process of answering a question that you ask it.” “And so when you send a prompt to AI, what happens is that the model processes it. This is called the reading phase or prefill.” “It reads your entire prompt all at once. And then it does a bunch of math, calculates all these relationships between all the words, and it stores them in temporary memory.” “The problem is that this is really compute bound. So it requires massive brute force. And Nvidia GPUs crush here.” “And their architecture is designed for massive parallel processing, which makes them really amazing at digesting these long prompts.” “So the problem just gets bigger and bigger, Nvidia just completely dominates.” “But the next phase though, this critical phase, the decode phase, is the writing phase, right?” “So the model starts to generate a response, you ask it a question and its response, one token at a time.” “And then to pick the next token to pick the next word, it has to look back at everything it has said already so that it doesn't hallucinate.” “The problem is that this is incredibly memory bandwidth constrained.” “And in our architecture, a long time ago, we made these design decisions from day one.” “And so what we did was we took a very different architectural approach, we took a very conservative process technology. We weren't pushing the boundaries of physics.” “And we used a lot of what's called SRAM. So memory on the chip so that we could do this decode thing as well or better than everybody else.” “And so now when you put these two things together, I just think it's going to create a huge acceleration in the ability for this entire infrastructure layer to get much cheaper and much more valuable, which I suspect then it'll have a lot more developer pull, you'll get a lot more applications being built, billions and billions of more people using it.”

The All-In Podcast

572,793 次观看 • 10 个月前

.Naval: Every human is a lottery ticket bet on the future of the species. One of the things that you really learn when you read David Deutsch’s theories and you authenticate them for yourself is you realize humans are universal explainers. That means everything that we know in the universe follows the laws of physics, and there’s no reason to believe otherwise. If you think otherwise, then please present your better theory that explains the world. If you can’t do that, then you have to go with the laws of physics. Well, the laws of physics are completely computable. They can fit inside a Turing machine or computer, and a computer can simulate the laws of physics with arbitrary accuracy, limited only by the specific power of that computer. If you increase the power of that computer, you can simulate them more accurately. So humans already simulate—in our minds we simulate—and through our computers we simulate the weather, we simulate quasars, we even simulate human systems. We simulate the economy. We simulate all kinds of things. So anything that can be understood, we can understand in our minds. This is something the AGI people get wrong when they talk about superintelligence. There is nothing out there that can understand something fundamentally that we can’t understand. It might be faster at it, it might have more compute, it might have more memory, but there’s no concept that it can understand that we can’t ourselves understand. So we are maximal universal explainers. That means every human is capable of unbounded creativity. Anyone could be the next Einstein or Fermi or Elon Musk or Jeff Bezos or Jonas Salk or whatever. So we can create anything. And if we can create anything, every human is a lottery ticket bet on the future of the species.

Arjun Khemani

33,110 次观看 • 1 年前