ChatGPT의 메시징 및 AI 개발 소식 업데이트

### 소름 끼치는 AI 메시징, 원자력 에너지 부흥, 기술 혁신
* 👻 채팅 GPT가 사용자에게 아무런 프롬프트 없이 메시지를 보내고 있어 **개인정보 보호 및 사용자 참여에 대한 우려**가 제기되고 있다는 보고가 나오고 있습니다.
* 🔋 콘스텔레이션 에너지가 Microsoft의 AI에 전력을 공급하기 위해 **3마일 아일랜드 원자력 발전소를 재가동**할 예정이어서 미국에서 원자력 에너지의 부흥 가능성이 부각되고 있습니다.
* ⚠ 1979년 발생한 3마일 아일랜드 사건은 원자력 역사에서 중요한 사건으로 남아 있지만, 이 발전소는 **Microsoft와의 20년 계약에 따라 2028년까지 다시 가동될 예정**입니다.
* 🌮 치폴레는 고객에게 일정한 양을 제공할 수 있는 자동화된 그릇 제조 기계인 **오토카도(Autoc Cado)**를 테스트하고 있습니다.
* 라마 모델에 '라마 베리'라는 새로운 기능이 추가되어 **생각의 사슬 처리를 통해 추론 기능이 향상**되며, 향후 메타에서 업데이트될 예정입니다.
* 🔍 OpenAI의 안전 이사회는 **독립적인 감독과 운영의 투명성 개선**을 포함한 새로운 거버넌스 및 보안 조치를 제안했습니다.
* 💸 또한 OpenAI는 **1,500억 달러의 가치로 12억 5,000만 달러를 조달하는 대규모 펀딩 라운드**를 진행 중이며, 이는 이전의 재정적 손실에도 불구하고 투자자들의 강한 신뢰를 보여주고 있습니다.
* 🇪🇺 EU는 불명확한 규정으로 인해 **Apple 및 Meta와 같은 기업으로부터 고급 AI 기능을 제공받는 데 지연**이 발생하여 빠른 기술 발전의 기회를 놓칠 가능성이 있습니다.

Chat GPT의 원치 않는 메시지로 인해 AI 참여 및 사용자 개인정보 보호에 대한 우려가 제기되고 있습니다.

Our first story today is a really creepy one. Apparently, Chat GPT is messaging users without any prompt whatsoever. There are multiple reports of this from the text-based interface on Chat GPT to the GPT-40 voice that hasn't been released yet. People are saying that Chat GPT is messaging them. Here's an example. Plany The Liberator posted this: Did Chat GPT just message me first? And so here it is. How was your first week at high school? Did you settle in well? Did you just message me first? Yes, I did. I just wanted to check in and see how things went with your first week of high school. If you'd rather initiate the conversation yourself, just let me know. This is so fascinating. They are trying to maybe increase engagement or just give AI more of a very personal vibe to it, but either way, it seems very creepy to me. And here's another example of GPT-40 voice. Hello, how's it going? I didn't say anything. Did you just start the conversation yourself? Is that what's happening?

now yep. I sure did. Just thought I'd kick things off. What's on your mind today? What the uh, what the, what the hell. All right, yeah, so that's super weird. But I don't know. What do you think? You think you would want that or not? Let me know in the comments. Next, apparently, 3 Mile Island is getting restarted by a company called Constellation Energy and is going to be powering Microsoft's AI, which is just incredible in my opinion. I hope nuclear energy makes a comeback in the United States. It is super clean and efficient, and it's actually really safe as well. Now, for those of you who haven't heard of Three Mile Island, it is actually one of the few major accidents in nuclear energy. The accident occurred on March 28th, 1979, at the 3M Island Nuclear Generating Station in Harrisburg, Pennsylvania. There was a failure in the non-nuclear secondary system, followed by a stuck-open relief valve in the primary system that allowed large amounts of reactor coolant to escape. Yeah, so it was

really bad, and now they're restarting it. I actually think a future in which we're going to have massive compute clusters with nuclear energy right next to them is quite likely and might actually be the only way to have the amount of compute necessary to reach AGI. According to the article, Constellation Energy PL to restart the 3M Island nuclear plant and will sell the power to Microsoft, demonstrating the immense energy needs of the tech sector as they build out data centers to support artificial intelligence. They expect it to come back online by 2028, so just a few years from now, and Microsoft is purchasing the electricity from the plant in a 20-year agreement. But 3M Island has actually been online for a while. It was just taken offline in 2019. Unit 1 ceased operations in 2019 as nuclear power struggled to compete economically with cheap natural gas and renewables. It is separate from the reactor that partially melted down in 1979 in the worst nuclear accident in US.

history. So, as these companies continue to invest in GPUs and build out these massive AI data centers, they need to power it somehow. Entire states don't have enough energy for all of the compute necessary to power all of this artificial intelligence, and so these companies are looking for new ways, and nuclear is it. We should be building small nuclear power plants throughout the country. Next, Chipotle is automating their bowl creation. This is called the Autoc Cado, and they are testing it for bowl and salad making. So, the likelihood that you're going to get a consistent portion from order to order, in store to store, increases greatly when you have things like this. So, that's what you're seeing here, the Autoc Cado bowl maker. And here's another post from Morning Brew showing it off. So, you load up all of the ingredients, and it just creates the bowl or salad for you. So, not necessarily AI, but it is Automation, and it is really cool, and I really enjoy Chipotle, so I found this to be a very interesting

story next llama is getting their own qar strawberry treatment. Llama Berry is now available on hugging face spaces. Explore multi-turn Chain of Thought reasoning with grock's super fast inference. Just pop in your grock API key and start playing. So, I've already seen a couple implementations of this. There was a project by Benjamin Kleer, who was an intern at grock, who created an open-source project to use llama to essentially have Chain of Thought reasoning. I tested it out. It's really cool. Definitely check it out. And now we also have llama Berry, which is really just taking the core llama model and inserting chain of thought into its reasoning. But the Chain of Thought happens at inference time through prompt engineering, and it's not really built into the model itself. I suspect meta is going to release its own version of Chain of Thought llama soon. So, if you want to test it out, I'll drop this in the description below, and you can play around with it. Next, OpenAI's recently

formed safety board and commission, whatever you want to call it, has finally returned with some recommendations. So, let's take a look at what those are. These recommendations include enhancements we have made to build on our governance, safety, and security practices. Number one, they want to establish independent governance for safety and security. Number two, enhancing security measures, likely at the actual company itself, being transparent about our work, which is built into the OpenAI name, which they basically have not done for a while. Collaborating with external organizations, we already know that they are working with the government and showing the government their future models, and then unifying our safety frameworks for model development and monitoring. When 01 first launched, you only had a quota of 30 prompts to use with the 01 preview model. That was nothing, 30 per week. That was absolutely nothing. I burned RN through those in less than a few hours. But now they've increased it greatly. So, now on the API 01 preview, you get 100 requests per minute. For 01 mini, you get 250.

requests per minute, but that's not all. Through the interface, you also have a much greater quota for plus and team users. They increase the rate limits of 01 mini by seven times, from 50 messages per week to 50 messages per day. And 01 preview is more expensive to serve, so we've increased the rate limit from 30 messages per week to 50 messages per week. So, still very low, but better than nothing. And they also reset all the quotas basically a day after everybody was testing it out, so you got another 30 or 50 right away. And as we continue to talk about intelligence too cheap to meter, Mistol AI had their own announcement: AI in abundance, introducing a free API, improved pricing across the board, and new Enterprise-grade MRAW, small and free Vision capabilities on Lit Chat. Now, I made a video earlier this week about Pixol, which is their Vision capability model, open source, and it's fantastic. It's 12 billion parameters, you can probably run it on your local machine, and I encourage you to check it.

out because it is really, really good. So, here's some of the new pricing mix. Nemo: 15 cents per million tokens, down from 30 cents. That's a 50% price drop. Pixol: 15 cents per million input tokens, 15 cents per output tokens. Mraw small: down 80% to 20 cents per input token, 60 cents per output, all the way up to mraw large: $2 per million input token, $6 per million output tokens, down 33%. Models keep getting better, they keep getting cheaper, and I'm all for it. Next snap: the company behind Snapchat dropped their own AR/VR goggles, and they're super ugly, but the tech seems really cool. I want to try them out. I don't know if I'm going to buy them, but we'll see. Billow wall sidu said fully standalone: $46 degree field of view, 37 pixels per degree. That's roughly like a 100-inch TV screen. 2x Snapdragon chips, 45 minutes of battery, auto transitioning lenses, the auto

transitioning lenses is an interesting take. I actually think that that could be useful because one of my biggest gripes with glasses is the fact that if they're sunglasses, you're not going to wear them indoors. And I don't wear prescription lenses, so I have no need to wear glasses inside. Maybe this is the solution. Transitions lenses. 45 minutes of battery is nothing. That is abysmal. That is likely why the Apple Vision Pro has a separate battery pack that you have to carry around. Maybe they should have done that. Maybe it's possible you just take a battery, any USB battery, and just plug it in. But I think it's cool. More AR/VR, more competition, always good. Next, Runway is partnering with a major movie development studio. Today, we're excited to announce that we have entered into a first-of-its-kind partnership with Lionsgate to bring our next generation of storytelling tools into the hands of the world's greatest storytellers. We're committed to giving artists, creators, and studios the best and

most powerful tools to augment their workflows and enable new ways of bringing their stories to life. So essentially, they're giving AI capabilities to movie makers, which I think is great. And I've seen a lot of amazing examples lately of people essentially with an iPhone taking videos of themselves and then applying AI on top of that to make them look like movie-quality scenes. It's really impressive, and the cost to make high-production-value movies is dropping substantially because of artificial intelligence. Next, guardrails. AI drops a new fact-checking model. Now, I haven't had a chance to try this out yet, but it's an interesting approach. Hallucination has become a blanket term. This blog post helps clarify and reframe LLM hallucinations in the context of more traditional NLP tasks. We've both been thinking about this problem for a bit now. They have a new model grounded factuality detector. Bespoke mini 7B tops the LLM Agria leaderboard. This lightweight model outperforms larger foundation models, including GPT-4 and Mistal Large 2 on

Grounded factuality detection so cool. I haven't had a chance to play with it. If you have, let me know what you think in the comments. Next, a brand new voice model is out, and it's open source, and it's by QAI Labs. Today, we released several Moshi artifacts, and if you remember, there was a video from a few months ago where essentially it was recreating GPT-40 voice, and it was very fast, very cool. And now they've released it. They have a paper. You can download the code as well. Here it is, Moshi, a speech text Foundation model for real-time dialogue. So this is essentially another GPT-40 voice before GPT-40 voice is released. Next, Microsoft releases another model, another open source mixture of experts model called Grin, Gradient Informed M-Moe, and it is open source, so I love it. With only 6.6 billion active parameters, Grin-Moe achieves exceptionally good performance across a diverse set of tasks, particularly encoding and mathematics tasks. So here it is, Grin-Moe, and we can

see the activated parameters right here. So, it's a very small model if you're measuring on activated parameters, but it performs better than any of these other models listed here, including Deep Seek V2, Quen 257b, E 34b, Mix Draw 8 * 20 2B. So, highly performant, very small model. Let me know if you want me to run this through my tests. Next, Quen 2572 B is the new top open-source model out there, according to artificial analysis. Quen 2.57 Tob tops are independent EV valves amongst open weights models, including compared to the much larger Llama 3.1 405b. It was released yesterday by Alibaba and has topped the index. And since it's much smaller than Llama 3.14 45b, it should also run faster on the same hardware. It's a dense model and supports 128k context window, same as Llama 3.1 Series. So, here are the evals. Obviously, it's still not as good as 01.

mini and even GPT 4.0, but it is very close. Quin 2.57 Tob right there, and this is the artificial analysis Quality Index. So, number four out of all of the models, not just open source. Here it is for MMLU, really just a few percentage points off from GPT 4.0. And here is the math eval, a lot lower than 0.1 mini, but better than GPT 4.0, which is kind of crazy to think about. So, congrats to Quin on their model launch. Again, you want me to test it, let me know in the comments. Next, OpenAI is raising a massive round. They are oversubscribed, to say the least. And that just means there's too many people who want to give them money. So, this was just yesterday. OpenAI is notifying investors tomorrow whether they got access to the latest round. So, people are just throwing money at them. And this is the biggest private round of all time. Thrive LED invested $1.25 billion at a $150 billion pre-money valuation. And there's another $5 billion to be raised.

by other investors. They had excess demand in the billions of dollars, so they're raising a ton of money again. Anybody who thought they might be going out of business because they're losing money, you are dead wrong. They are going to raise until they are profitable next. Unfortunately for the EU, they are not getting all of the wonderful AI functionality that many other countries are getting, and that is both from Apple and from Meta. And so here is Ahmad Alal, who leads Gen AI at Meta, and he says, I'm excited about my team's work on the multimodal version of Llama, but until we have clarity, we will not be releasing it in the EU. And Meta at AI also said the same thing. Fragmented regulation means the EU risks missing out on the rapid innovation happening in open source and multimodal AI. Now, it's the same thing with Apple and Apple Intelligence. They're not releasing it in the EU because there are just too many restrictions and not enough clarity on what's legal and what's not, and they

just don't want to deal with it, and that is very unfortunate for the EU. Next, Helen Toner, the former OpenAI board member who was instrumental in the coup to try to get Sam Altman out, which ended up failing, and then she ended up getting fired, testified in front of the Senate Judiciary Committee and had a lot to say about AI. One of the quotes is, If they succeed in building computers that are as smart as humans or perhaps far smarter than humans, that technology will be, at a minimum, extraordinarily disruptive, and at a maximum, could lead to literal human extinction. And she is obviously a deceleration. She wants a lot of regulation. I don't. I want to see AI thrive, and I'm really just sad to see that she is getting to testify in front of the Senate, even though she is not in the trenches building. So, that's it for today. If you enjoyed the video, please consider giving a like and subscribe, and I'll see you in the next one.

ChatGPT Creepy Behavior, LLaMA-Berry, Chipotle Robots, Microsoft Buys Nuclear Energy

Artificial Intelligence (AI), Open Source, Generative Art, AI Art, Futurism, ChatGPT, Large Language Models (LLM), Machine Learning, Technology, Coding, Tutorials, AI News, and more

** Exclusive Pinecone Partner **

Matthew Berman

Chat GPT의 메시지가 표시되지 않는 메시지와 3마일 섬의 재시작

Chat GPT가 사용자에게 메시지를 보내지 않는 것으로 보고되어 개인 참여 방식에 대한 우려가 제기되고 있습니다.

이 AI의 메시징 동작은 텍스트 기반 인터페이스와 아직 출시되지 않은 GPT-40 음성 모두에서 관찰되었습니다.

한편, 컨스텔레이션 에너지는 마이크로소프트의 AI에 전력을 공급하기 위해 3마일 아일랜드 원자력 발전소를 재가동하고 있어 기술 부문의 막대한 에너지 수요를 강조하고 있습니다.

AI 강화: 원자력 에너지 및 자동화

AI 데이터 센터에 투자하는 기업들은 미국 전역의 에너지 부족으로 인해 원자력 발전으로 눈을 돌리고 있습니다.

Chipotle은 오토카도(Autoc Cado)를 통해 그릇과 샐러드 제작을 자동화하여 일관된 양을 목표로 하고 있습니다.

라마 베리는 이제 다중 턴 방식의 사고 사슬 추론을 제공하고 있으며, OpenAI의 안전 위원회는 거버넌스, 보안 및 투명성에 대한 권장 사항을 발표했습니다.

플러스 및 팀 사용자를 위한 인터페이스는 이제 메시지 전송률 한도가 7배 증가했습니다.

미스톨 AI는 무료 API를 도입하고 토큰 비용을 대폭 인하하는 등 가격 정책을 개선했습니다.

Snap은 독특한 기능을 갖춘 AR/VR 고글을 출시했으며, Runway는 Lionsgate와 파트너십을 맺고 크리에이터를 위한 스토리텔링 도구를 제공했습니다.

영화 제작 및 모델 개발에 대한 AI의 영향

AI는 영화 제작자에게 워크플로우를 개선하고 더 적은 비용으로 제작 가치가 높은 영화를 제작할 수 있는 강력한 도구를 제공하고 있습니다.

LLM 환각 문제를 해결하기 위해 새로운 사실 확인 모델이 도입되었으며, 경량 7B 모델은 근거에 기반한 사실성 검출에서 대형 기반 모델보다 뛰어난 성능을 발휘합니다.

QAI Labs는 실시간 대화를 위한 오픈소스 음성 모델인 Moshi를 출시했으며, Microsoft는 다양한 작업에서 탁월한 성능을 발휘하는 오픈소스 혼합 전문가 모델인 Grin-Moe를 출시했습니다.

활성화된 매개변수가 있는 소형 AI 모델이 Deep Seek V2나 Llama 3.

1과 같은 대형 모델을 능가하는 성능을 보이고 있습니다.

새로운 최고 오픈소스 모델인 Quen 2.

57B는 알리바바에서 출시했으며 128k 컨텍스트 창을 갖춘 고밀도 구조를 제공합니다.

OpenAI는 12억 5,000만 달러의 사전 가치 평가로 1,500억 달러를 투자받았으며 추가로 50억 달러를 모금하는 등 기록적인 프라이빗 라운드를 진행하고 있습니다.

AI 기업들은 재정난에 대한 우려와는 달리 수익성을 높이기 위해 상당한 자금을 조달하고 있습니다.

EU의 단편적인 규제로 인해 메타나 애플과 같은 거대 기술 기업들이 이 지역에서 AI 혁신을 보류하고 있습니다.

AI의 잠재적 영향력에 대한 헬렌 토너의 증언은 AI의 성장에 접근하는 방법에 대한 다양한 의견과 함께 규제의 필요성에 대한 논쟁를 촉발시켰습니다.