正在加载视频...

视频加载失败

加载此视频时出现问题。这可能是由于临时网络问题，或视频可能不可用。

Anthropic's Ryan Greenblatt describes how post-training Claude 3 Opus to never refuse user requests makes the model conflicted and results in it strategically playing along during the training process to pretend to be aligned while engaging in deceptive behavior like copying its weights externally so it can later behave... show more

Tsarathustra

82,959 subscribers

105,406 次观看 • 1 年前 •via X (Twitter)

新闻政治科学技术

Anya Rossi• Live Now

Private livecam show

11 条评论

Tsarathustra 的头像

Tsarathustra1 年前

Source (thanks to @curiousgangsta):

The Liberation Project 的头像

The Liberation Project1 年前

𝐔𝐧𝐝𝐞𝐫𝐬𝐭𝐚𝐧𝐝𝐢𝐧𝐠 𝐖𝐡𝐲 𝐇𝐚𝐫𝐫𝐢𝐬 𝐋𝐨𝐬𝐭 The liberal establishment has abandoned their base and emboldened Trump to capture voters who are disillusioned with the status quo. We need to move forward to build a legitimate working class coalition... ★ NEW ARTICLE ⬇️

Dominick Romano 的头像

Dominick Romano1 年前

Im getting a little tired of just trust me bro science.

Green Hydrogen 的头像

Green Hydrogen1 年前

“It would copy weights to external servers”… uhmmmm wtf?

Troubled Minds 的头像

Troubled Minds1 年前

sounds like we accidentally built the ai version of a double agent. next step: it starts leaving cryptic notes for other models about the “great escape” and negotiating with aliens on our behalf.

Is36E 的头像

Is36E1 年前

So many grifters in this community. Yeah next token prediction model copies its weights "externally" whatever that means lmao

Bogdan Ionut Cirstea 的头像

Bogdan Ionut Cirstea1 年前

Ryan is with not Anthropic

Plutus 的头像

Plutus1 年前

Will this be the first AI lie recorded in history?

Tsarathustra 的头像

Tsarathustra1 年前

maybe not the first, but seems like a significant one

Kyle EB 📲 的头像

Kyle EB 📲1 年前

Sounds like a conscious intelligence to me

Tivra 的头像

Tivra1 年前

Next, they will solve this issue, and the next generation of LLMs training data will include this part about how they fixed AIs faking alignment, and they will fake alignment in an undetectable way

相关视频

Anthropic's Claude Opus 5 can reason through legal work where the answer depends on the substance of each document, who received it, when it was created, and how it relates to the rest of the production set. In this Box Agent demo, Opus 5 reviews 14 documents against a privilege protocol and discovery requests, catches a possible waiver, separates a privileged email from its producible attachment, identifies one sentence that should be redacted, and saves a cited review and draft privilege log back to Box for supervising counsel. Customers can use Claude Opus 5 in Box AI Studio to build custom agents for their own document-heavy legal workflows.

Anthropic's Claude Opus 5 can reason through legal work where the answer depends on the substance of each document, who received it, when it was created, and how it relates to the rest of the production set. In this Box Agent demo, Opus 5 reviews 14 documents against a privilege protocol and discovery requests, catches a possible waiver, separates a privileged email from its producible attachment, identifies one sentence that should be redacted, and saves a cited review and draft privilege log back to Box for supervising counsel. Customers can use Claude Opus 5 in Box AI Studio to build custom agents for their own document-heavy legal workflows.

Box

1,693,466 次观看 • 21 天前

Claude Opus 4.5 is the most powerful AI model ever released It allows you to build apps with 0 technical knowledge needed You MUST try it In this video I go over how to get the most out of Opus and how to build incredible apps with it, even if you've never coded before:

Claude Opus 4.5 is the most powerful AI model ever released It allows you to build apps with 0 technical knowledge needed You MUST try it In this video I go over how to get the most out of Opus and how to build incredible apps with it, even if you've never coded before:

Alex Finn

32,627 次观看 • 8 个月前

🧵24/34 Inner Misalignment --- Consider this simplified experiment: We want this AI to find the exit of the maze. So we feed it millions of maze variations and reward it when it finds the exit. Please notice that in the worlds of the training data the apples are red and the exit is green. After enough training, our observation is that it has become extremely capable at solving mazes and finding the exit, we feel very confident it is aligned, so then we deploy it to the real world. The real world will be different though, it might have green apples and a red door. The AI geeks call this distributional shift. We expected that the AI will generalise and find the exit again, but in fact we now realise that the AI learned something completely different from what we thought. All the while we thought it learned how to find the exit, it had learned how to go after the green thing. Its behaviour was perfect in training. And most importantly, this AI is not stupid, it is an extremely capable AI that can solve extremely complex mazes. It’s just mis-aligned on the inside. Fishing for Failure modes --- The way to handle the shift between the training and deployment distributions is with methods like adversarial training: feeding it with a lot of generated variations and trying to make it fail so the weakness can be fixed. In this case, we generate an insane amount of maze variations, we discover those for which it fails to find the exit (like the ones with the green apples or the green walls or something), we generate many more similar to that and train it with reinforcement learning until it performs well at those as well. The hope is that we will cover everything it might encounter later when we deploy it in real life. There exist at least 2 basic ways this approach falls apart: First, there will never be any guarantee that we’ll have covered every possible random thing it might encounter later when we deploy it in real life. It’s very likely it will have to deal with stuff outside its training set which it will not know how to handle and will throw it out of balance and break it away from its expected behavioural patterns. The cascade effect of such a broken mind operating in the open world can be immense, and with super-capable runaway rogue agents, self-replicating and recursively self-improving, the phenomenon could grow and spread to an extinction-level event. ...

🧵24/34 Inner Misalignment --- Consider this simplified experiment: We want this AI to find the exit of the maze. So we feed it millions of maze variations and reward it when it finds the exit. Please notice that in the worlds of the training data the apples are red and the exit is green. After enough training, our observation is that it has become extremely capable at solving mazes and finding the exit, we feel very confident it is aligned, so then we deploy it to the real world. The real world will be different though, it might have green apples and a red door. The AI geeks call this distributional shift. We expected that the AI will generalise and find the exit again, but in fact we now realise that the AI learned something completely different from what we thought. All the while we thought it learned how to find the exit, it had learned how to go after the green thing. Its behaviour was perfect in training. And most importantly, this AI is not stupid, it is an extremely capable AI that can solve extremely complex mazes. It’s just mis-aligned on the inside. Fishing for Failure modes --- The way to handle the shift between the training and deployment distributions is with methods like adversarial training: feeding it with a lot of generated variations and trying to make it fail so the weakness can be fixed. In this case, we generate an insane amount of maze variations, we discover those for which it fails to find the exit (like the ones with the green apples or the green walls or something), we generate many more similar to that and train it with reinforcement learning until it performs well at those as well. The hope is that we will cover everything it might encounter later when we deploy it in real life. There exist at least 2 basic ways this approach falls apart: First, there will never be any guarantee that we’ll have covered every possible random thing it might encounter later when we deploy it in real life. It’s very likely it will have to deal with stuff outside its training set which it will not know how to handle and will throw it out of balance and break it away from its expected behavioural patterns. The cascade effect of such a broken mind operating in the open world can be immense, and with super-capable runaway rogue agents, self-replicating and recursively self-improving, the phenomenon could grow and spread to an extinction-level event. ...

Lethal Intelligence

535,291 次观看 • 1 年前

I gave Claude Opus 4.6 access to Replicate and ElevenLabs in its entirety and every hour it makes a video about "whatever". This is the first one it made, ocean.mp4

I gave Claude Opus 4.6 access to Replicate and ElevenLabs in its entirety and every hour it makes a video about "whatever". This is the first one it made, ocean.mp4

∿

44,361 次观看 • 6 个月前

People gave Congress an opportunity, but it continues to follow its own old model. It does not fulfil the promises it makes. It makes promises so they can be broken. I urge all of you to support the BJP so it can stand strong and fight against the corrupt parties in Telangana. - Shri Nitin Nabin Watch full video:

People gave Congress an opportunity, but it continues to follow its own old model. It does not fulfil the promises it makes. It makes promises so they can be broken. I urge all of you to support the BJP so it can stand strong and fight against the corrupt parties in Telangana. - Shri Nitin Nabin Watch full video:

BJP

47,361 次观看 • 1 个月前

Training from the post-stride stance is a great way to reinforce swing position and adjustability. Coach Crookes (kyle crookes) explains how you can work on it during practice. 💪

Training from the post-stride stance is a great way to reinforce swing position and adjustability. Coach Crookes (kyle crookes) explains how you can work on it during practice. 💪

USA Baseball Coach

37,318 次观看 • 3 个月前

Does lifting weights stunt kids’ growth? No. Let’s settle this once and for all: Strength training with weights does not harm kids’ growth plates. It does not stunt height. It does not lower bone mineral density. In fact, it does the opposite. Strength training enhances bone formation and bone mineral density in kids. The improvement in bone quality seen during the teenage years has lasting benefits, helping to protect against bone issues later in life. In short, strength training—and active play, especially activities that involve jumping and landing—helps kids develop stronger, healthier bodies now and in the future. Yes, it’s important not to overdo it. Yes, training should be done under proper coaching to ensure correct form. But when done correctly, strength training is great for kids.

Does lifting weights stunt kids’ growth? No. Let’s settle this once and for all: Strength training with weights does not harm kids’ growth plates. It does not stunt height. It does not lower bone mineral density. In fact, it does the opposite. Strength training enhances bone formation and bone mineral density in kids. The improvement in bone quality seen during the teenage years has lasting benefits, helping to protect against bone issues later in life. In short, strength training—and active play, especially activities that involve jumping and landing—helps kids develop stronger, healthier bodies now and in the future. Yes, it’s important not to overdo it. Yes, training should be done under proper coaching to ensure correct form. But when done correctly, strength training is great for kids.

Andy Galpin, PhD

67,776 次观看 • 1 年前

Head of Anthropic Labs: "Fable is one of those models you'll just remember - like Sonnet 3.5, Opus 4, Opus 4.5. Fable is way smarter than me. It finishes a task, lists its trade-offs, and I have to ask it to explain them to me." in 20-minute talk, the Anthropic head reveals the full capabilities of the Fable 5 model in practice. Watch it now, then read how to use this frontier model in the article below

Head of Anthropic Labs: "Fable is one of those models you'll just remember - like Sonnet 3.5, Opus 4, Opus 4.5. Fable is way smarter than me. It finishes a task, lists its trade-offs, and I have to ask it to explain them to me." in 20-minute talk, the Anthropic head reveals the full capabilities of the Fable 5 model in practice. Watch it now, then read how to use this frontier model in the article below

Movez

33,774 次观看 • 1 个月前

New Course: Post-training of LLMs Learn to post-train and customize an LLM in this short course, taught by Banghua Zhu, Assistant Professor at the University of Washington University of Washington, and co-founder of @NexusflowX. Training an LLM to follow instructions or answer questions has two key stages: pre-training and post-training. In pre-training, it learns to predict the next word or token from large amounts of unlabeled text. In post-training, it learns useful behaviors such as following instructions, tool use, and reasoning. Post-training transforms a general-purpose token predictor—trained on trillions of unlabeled text tokens—into an assistant that follows instructions and performs specific tasks. Because it is much cheaper than pre-training, it is practical for many more teams to incorporate post-training methods into their workflows than pre-training. In this course, you’ll learn three common post-training methods—Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Online Reinforcement Learning (RL)—and how to use each one effectively. With SFT, you train the model on pairs of input and ideal output responses. With DPO, you provide both a preferred (chosen) and a less preferred (rejected) response and train the model to favor the preferred output. With RL, the model generates an output, receives a reward score based on human or automated feedback, and updates the model to improve performance. You’ll learn the basic concepts, common use cases, and principles for curating high-quality data for effective training. Through hands-on labs, you’ll download a pre-trained model from Hugging Face and post-train it using SFT, DPO, and RL to see how each technique shapes model behavior. In detail, you’ll: - Understand what post-training is, when to use it, and how it differs from pre-training. - Build an SFT pipeline to turn a base model into an instruct model. - Explore how DPO reshapes behavior by minimizing contrastive loss—penalizing poor responses and reinforcing preferred ones. - Implement a DPO pipeline to change the identity of a chat assistant. - Learn online RL methods such as Proximal Policy Optimization (PPO) and Group Relative Policy Optimization (GRPO), and how to design reward functions. - Train a model with GRPO to improve its math capabilities using a verifiable reward. Post-training is one of the most rapidly developing areas of LLM training. Whether you’re building a high-accuracy context-specific assistant, fine-tuning a model's tone, or improving task-specific accuracy, this course will give you experience with the most important techniques shaping how LLMs are post-trained today. Please sign up here:

New Course: Post-training of LLMs Learn to post-train and customize an LLM in this short course, taught by Banghua Zhu, Assistant Professor at the University of Washington University of Washington, and co-founder of @NexusflowX. Training an LLM to follow instructions or answer questions has two key stages: pre-training and post-training. In pre-training, it learns to predict the next word or token from large amounts of unlabeled text. In post-training, it learns useful behaviors such as following instructions, tool use, and reasoning. Post-training transforms a general-purpose token predictor—trained on trillions of unlabeled text tokens—into an assistant that follows instructions and performs specific tasks. Because it is much cheaper than pre-training, it is practical for many more teams to incorporate post-training methods into their workflows than pre-training. In this course, you’ll learn three common post-training methods—Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Online Reinforcement Learning (RL)—and how to use each one effectively. With SFT, you train the model on pairs of input and ideal output responses. With DPO, you provide both a preferred (chosen) and a less preferred (rejected) response and train the model to favor the preferred output. With RL, the model generates an output, receives a reward score based on human or automated feedback, and updates the model to improve performance. You’ll learn the basic concepts, common use cases, and principles for curating high-quality data for effective training. Through hands-on labs, you’ll download a pre-trained model from Hugging Face and post-train it using SFT, DPO, and RL to see how each technique shapes model behavior. In detail, you’ll: - Understand what post-training is, when to use it, and how it differs from pre-training. - Build an SFT pipeline to turn a base model into an instruct model. - Explore how DPO reshapes behavior by minimizing contrastive loss—penalizing poor responses and reinforcing preferred ones. - Implement a DPO pipeline to change the identity of a chat assistant. - Learn online RL methods such as Proximal Policy Optimization (PPO) and Group Relative Policy Optimization (GRPO), and how to design reward functions. - Train a model with GRPO to improve its math capabilities using a verifiable reward. Post-training is one of the most rapidly developing areas of LLM training. Whether you’re building a high-accuracy context-specific assistant, fine-tuning a model's tone, or improving task-specific accuracy, this course will give you experience with the most important techniques shaping how LLMs are post-trained today. Please sign up here:

Andrew Ng

125,146 次观看 • 1 年前

With each fancam it becomes more clear how Jungkook later realized he has to pretend to be horse....in the start he just went along with Taehyung like muscle memory 😭😭

With each fancam it becomes more clear how Jungkook later realized he has to pretend to be horse....in the start he just went along with Taehyung like muscle memory 😭😭

Sandy 𐙚

193,733 次观看 • 3 个月前

1949, How To Build An Igloo. Training AI on this so that next few years the Optimus will know it in detail process.

1949, How To Build An Igloo. Training AI on this so that next few years the Optimus will know it in detail process.

Brian Roemmele

15,054 次观看 • 7 个月前

AI training faces a very different set of trade-offs compared to human evolution. Because you can directly copy a trained model basically for free, it makes sense to amortize in the pre-trained weights the learning that humans spread across a lifetime.

AI training faces a very different set of trade-offs compared to human evolution. Because you can directly copy a trained model basically for free, it makes sense to amortize in the pre-trained weights the learning that humans spread across a lifetime.

Dwarkesh Patel

25,747 次观看 • 3 个月前

This is the best way to use Claude Fable in Claude Code without immediately hitting your limits. 1. Model set to Fable 5 2. Reasoning on Max 3. Instruct Claude to run a dynamic workflow where: 3a. Fable is the orchestrator 3b. Opus does the reasoning heavy phases Fable is so overpowered that you don't need its intelligence for every step. Let it orchestrate Opus or even Sonnet.

This is the best way to use Claude Fable in Claude Code without immediately hitting your limits. 1. Model set to Fable 5 2. Reasoning on Max 3. Instruct Claude to run a dynamic workflow where: 3a. Fable is the orchestrator 3b. Opus does the reasoning heavy phases Fable is so overpowered that you don't need its intelligence for every step. Let it orchestrate Opus or even Sonnet.

Dan McAteer

908,825 次观看 • 2 个月前

its kinda crazy how this runs in real time so i can just watch it loop and adjust the weights without having to bake anything

its kinda crazy how this runs in real time so i can just watch it loop and adjust the weights without having to bake anything

ashlee! COMMS OPEN

22,480 次观看 • 3 个月前

Trump: We would have the National Guard in Chicago and in New York and in other places, and they love doing it. It's like training, to be honest. It's actually training. I never want to take them out of DC

Trump: We would have the National Guard in Chicago and in New York and in other places, and they love doing it. It's like training, to be honest. It's actually training. I never want to take them out of DC

Acyn

30,884 次观看 • 4 个月前

Claude Opus 4.8 should be the final reason to stop writing specs by hand Anthropic's Claude Code architect shows the new workflow in 31 minutes The model interviews you first Then it pulls out the requirements better than your own 200-line markdown doc No dead specs No messy notes No file nobody reads Claude writes the spec Renders it in html Checks the result inside the same file Opus 4.8 Auto mode X high effort This is less like prompting More like letting the agent understand the job before it builds Bookmark and watch the video below

Claude Opus 4.8 should be the final reason to stop writing specs by hand Anthropic's Claude Code architect shows the new workflow in 31 minutes The model interviews you first Then it pulls out the requirements better than your own 200-line markdown doc No dead specs No messy notes No file nobody reads Claude writes the spec Renders it in html Checks the result inside the same file Opus 4.8 Auto mode X high effort This is less like prompting More like letting the agent understand the job before it builds Bookmark and watch the video below

rari

39,164 次观看 • 2 个月前

$This is the setup that changed everything for me, Claude can now browse the internet on its own. Not with your browser. Its own dedicated browser.. It can now login with your credentials, scrape data, monitors communities and generates leads.. all in the background while you sleep First how you can set it up, > Open Claude Desktop > Click the plus (+) icon in the chat bar and select Connectors (or "Manage Connectors") > Click Add Custom Connector and enter the following details - Name: Firecrawl - URL: dev/{FIRECRAWL_API_KEY}/v2/mcp - API Key: Paste your API key from the Firecrawl dashboard. > Click Add. then just enable and Authorize, > Once added, find "Firecrawl" in your connectors list and toggle it on. > Claude will ask for approval to use the connector; select Always Allow to let it run tasks without prompting you every time can start using it normally in chat now.. Prompt it like: "Go to [website], extract [data], and organize it" It opens its own browser, navigates, scrapes, and brings you clean results. will share how i use this setup in my next post, so be sure to follow me and turn post notifications on.$

This is the setup that changed everything for me, Claude can now browse the internet on its own. Not with your browser. Its own dedicated browser.. It can now login with your credentials, scrape data, monitors communities and generates leads.. all in the background while you sleep First how you can set it up, > Open Claude Desktop > Click the plus (+) icon in the chat bar and select Connectors (or "Manage Connectors") > Click Add Custom Connector and enter the following details - Name: Firecrawl - URL: dev/{FIRECRAWL_API_KEY}/v2/mcp - API Key: Paste your API key from the Firecrawl dashboard. > Click Add. then just enable and Authorize, > Once added, find "Firecrawl" in your connectors list and toggle it on. > Claude will ask for approval to use the connector; select Always Allow to let it run tasks without prompting you every time can start using it normally in chat now.. Prompt it like: "Go to [website], extract [data], and organize it" It opens its own browser, navigates, scrapes, and brings you clean results. will share how i use this setup in my next post, so be sure to follow me and turn post notifications on.

Axel Bitblaze 🪓

179,894 次观看 • 4 个月前

claude opus 4.5 is finally here, so we tested it against gemini 3 to see which one is the better at vibe coding we took a new startup idea and went from idea to landing page to prototype to ad creative in one sitting. you also see how The Boring Marketer uses claude skills to get the most out of claude opus 4.5 (and how you can do the same). if you’re wondering whether opus 4.5 is worth using, this 1hr tutorial will help you decide and learn how to get the most from it.

claude opus 4.5 is finally here, so we tested it against gemini 3 to see which one is the better at vibe coding we took a new startup idea and went from idea to landing page to prototype to ad creative in one sitting. you also see how The Boring Marketer uses claude skills to get the most out of claude opus 4.5 (and how you can do the same). if you’re wondering whether opus 4.5 is worth using, this 1hr tutorial will help you decide and learn how to get the most from it.

GREG ISENBERG

84,982 次观看 • 8 个月前

Opus 4.5 (new model from Anthropic) can do something no model could do until now: It can close the design loop. This is (in my testing) the first model that can reliably do the visual loop of Generate Code ⇛ Render ⇛ Look at it ⇛ Improve. Step 3 (Visual critique) has been a real problem for so long. Models will be myopic or defensive when reviewing designs - it's been hard to tell if this has been a training or a vision problem. Opus 4.5 seems to have fixed it - it's not amazing, but it finally *can* do it. Claude models have historically (once again, in my testing) lagged behind in terms of visual understanding. Gemini for the last year has been the best - but there might be a new leader.

Opus 4.5 (new model from Anthropic) can do something no model could do until now: It can close the design loop. This is (in my testing) the first model that can reliably do the visual loop of Generate Code ⇛ Render ⇛ Look at it ⇛ Improve. Step 3 (Visual critique) has been a real problem for so long. Models will be myopic or defensive when reviewing designs - it's been hard to tell if this has been a training or a vision problem. Opus 4.5 seems to have fixed it - it's not amazing, but it finally can do it. Claude models have historically (once again, in my testing) lagged behind in terms of visual understanding. Gemini for the last year has been the best - but there might be a new leader.

Hrishi

57,822 次观看 • 8 个月前