주간 AI 업데이트: 새로운 모델과 혁신 공개

### 금주의 AI 하이라이트 새로운 모델, 도구, 획기적인 기술
* 🍓 OpenAI의 새로운 모델인 QAR(GPT-5)에 대한 스트로베리 씨의 트위터 과대 광고는 대부분 정확하지 않음에도 불구하고 **밈을 생성**했습니다.
* 🤖 엘론 머스크의 AI 회사인 X에서 출시한 **Grock 2 베타 및 Grock 2 Mini**는 **고급 로직, 추론 및 텍스트-이미지 변환 기능**을 갖추고 있습니다.
* 🔍 최신 정보와 미디어가 풍부한 검색 결과를 제공하는 **잠재적인 Google 검색 엔진 경쟁자인 Search GPT**가 미리 공개되었습니다.
* 🧠 멀티온은 계획 수립 및 자가 치유 기능을 갖춘 새로운 AI 에이전트인 **에이전트 Q**를 도입하여 실제 업무에서 놀라운 성능 향상을 달성했습니다.
* 💻 코사인 도시는 **S 벤치에서 30%, U 라이트 벤치마크에서 50%**를 달성한 최첨단 소프트웨어 엔지니어링 모델인 **Genie**를 출시했습니다.
* 🔬 사카나 AI의 AI 과학자는 **완전히 자동화된 과학적 발견, 새로운 연구 아이디어 생성, 코드 작성, 실험 수행, 논문 동료 검토**를 약속합니다.
* 📊 OpenAI는 AI 모델의 소프트웨어 엔지니어링 능력을 보다 안정적으로 평가하기 위해 **사람이 검증한 sbench의 하위 집합인 sbench verified**를 도입했습니다.
* 🤖 대규모 언어 모델에 유용한 기능인 Claude를 통해 **다양한 AI 애플리케이션의 비용을 절감**하고 속도를 높여주는 **프롬프트 캐싱**을 출시했습니다.

스트로베리 씨의 급격한 트위터 성장세는 이번 주 AI 개발의 혼란스러운 분위기를 잘 드러냈습니다.

All right, this week was absolutely insane. We had so many awesome releases. We're going to talk about it today. I also just got done with a live stream where we went over a lot of these stories, and I had some special guests. You got to meet some of the people on the team, so be sure to check out the next live stream. I think we're going to be doing it every Friday at 10:00 a.m. Pacific. Our first story, let's start with Mr. Strawberry. I rule the world, Mo. He absolutely took over Twitter, AI Twitter that is, this week, and went from 3,000 followers to 33,000 followers, pretty much hyping the hell out of Strawberry, which is QAR, whatever you want to call it. It is the new logic and reasoning model, maybe GPT-5 from OpenAI. But here's the thing, he was wrong about most of it, and you know what? It was fun. I don't really blame him for putting out a lot of BS, but at the end of the day.

There were a lot of memes, and that's really all that matters, right? So I'm not going to talk too much about this, but yeah, he was definitely wrong. And next, Grock 2 Beta is released. I already made a video about this; I'll drop a link in the description below so you can check out a full video about it. Grock 2 Beta released, and here's the thing: you remember that model Sus Colr that showed up on lm.org as this Anonymous model that nobody really knew what it was? Well, now we know it was Grock 2. I already did a video testing Sus Colar, so be sure to check that out. I'll also drop that in the description below. So this is the new model release from X, which is Elon Musk's AI company. They get a lot of their training data from X itself. We are excited to release an early preview of Grock 2, a significant step forward from our previous model Grock 1.5, featuring frontier capabilities in chat, coding, and

reasoning at the same time. We are introducing Grock 2 Mini, a small but capable sibling of Grock 2. So, right now, the only one that's actually available live is Grock 2 Mini, not Grock 2. That should be coming soon. Grock 2 is incredibly good at logic and reasoning, and it has a text-to-image model powered by Flux. One. And as a quick reminder, Flux. One is a completely open-source text-to-image model, which has great quality on par with Mid Journey. And it was started by a group of folks who left Stable Diffusion to create a new company, Flux. And it says right here, Grock 2 and Grock 2 Mini are currently in beta on X. But that's not actually true. As you can see here, the only one you can get is Grock 2 Mini. But people have been going absolutely crazy with the uncensored nature of Grock 2, and more importantly, the text-to-image model. You can basically get anything except nudity. All

Right, let's look at some examples of the text images that are being created. Here's one of Kamala and Donald Trump. Kamala is pregnant, wearing a crown. I mean, some of these photos are just absurd. Here's one that started with a Gro image and then was animated with another tool. Here's one of Donald Trump driving what looks like the Mars rover, flying an American flag on Mars. There's a lot of Donald Trump ones here. Here's Donald Trump and Abe Lincoln together. Here is the sexiest family on Earth, according to Adriano. Okay, oh yeah, that is definitely a sexy dog. Everybody seemingly is making images of Donald Trump for some reason. Here's one as Darth Vader. Here's another Chewbacca and Han Solo Trump. Love it. And Grok 2 does not care about copyright. Here's Nuto. Here's the Flintstones, Simpsons, and this looks incredible. And I actually don't know who this character is, but that's the North Korean leader in the background. And here's another one of Peter Griffin from Family Guy. So as you

can see you can go absolutely nuts with grock to text the image. Check it out next. People started getting early access to search GPT, and I'm one of them. So we actually got Serge GPT. Let's give it a try. Tell me about the UFC fight this weekend. We get a bunch of source on the left, and then it writes out a response over here. We can give it a thumbs up or thumbs down, good link, bad link. And we also can get media from it, and it is up-to-date information. I've already switched my default search engine to B search GPT, and I'm not looking back. There is really no reason for me to go to Google anymore. I am sorry to say, but it is absolutely true. Google's search dominance is more than being threatened now. It is being dominated, and it's just a matter of time until perplexity and search GPT take over completely. I want answers. I don't want ads, and I don't want to have to look through 10 Blue Links. So I'm thinking about doing a full

review of search GPT, but I'm not sure it warrants a full video. But let me know what you want in the comments below. Next from multi on, they are releasing agent Q, research breakthrough for the next generation of AI agents with planning and self-healing capabilities. So apparently, this is kind of the next generation of agent. Now, it is seemingly specifically for consumers. This isn't for Enterprise, and it's closed source, so keep that in mind. But introducing agent Q. In recent years, the capabilities of large language models have transformed natural language processing and understanding, achieving remarkable milestones. Despite these advancements, LLMs face significant challenges in interactive environments, especially in tasks requiring multi-step reasoning like web navigation. So enter agent Q, a major milestone for agents combining search, self-critique, and reinforcement learning to create state-of-the-art autonomous web agents that can plan and self-heal. I definitely want to try it out. I haven't tried it yet. It looks really cool. But let's look at the actual benchmarks. So in real-world booking

experiments on Open Table multi-on agents drastically improved the zero-shot performance of the Lama 3 model from an 18% to 81%. That is astounding, just after 1 day of autonomous data collection, and further 9.5% with online search. Amazing. So, they have a whole paper here. Let me know if you want me to break down the paper in a separate video. I'm happy to do that. They used a lot of components of what strawberry and qar are supposed to be, so guided search with MCTS (that is Monte Carlo research), Ai self-critique, and DPO (direct preference optimization). Now, why is this important? Because a lot of people have been talking about strawberry and qar lately, and these are the elements that have supposedly made up qar. The multi-on team has attached themselves to Mr. Strawberry, I rule the world Mo, claiming that that is them. But who knows if that's actually true? Nobody actually knows who I rule the world is. But either way, looks like another cool agent that

we get to play around with, so I'm excited. Next, from cosign Dosh, we have the best software engineering model to date. Genie is the best software engineering model in the world, scoring a state-of-the-art score of 30% on S bench and 50% on U light. Now, what does that actually mean? This is a model that is trained to write code, and it writes code incredibly well. You give it a task, whatever that task is, and it will write the code, correct existing code, update existing code, whatever you need. And here are the benchmarks for it. So, we see cosign Genie at the top with a 30% on sweep bench. Now, the most popular coding agent to date has been cognition Devon, and this is the extrapolated version coming in at 14%, less than half of the score of cosign Genie. So, they really made a huge jump in performance. And if you look at cognition Devon down here, this is just another version. I think maybe the actual tested version versus the self-reported version. I'm not sure.

what the difference is, but this is 4%. So, as you can see, a much better performance than anything else on the market. It was trained with 21% JavaScript, 21% Python, 14% TypeScript, and so on. My favorite programming language, aside from Python, Ruby, coming in at a measly 3%. And here's the data mix: feature development 25%, bug fixing 20%, refactoring 15%, minor changes and chores 15%, test writing incredibly important, especially for AI code builders 15%, and writing documentation 10%. Genie was always designed to be a genetic as well, so if you haven't had a chance to check it out, please do. Also, join the waitlist. Next is possibly one of the most exciting stories of the week and almost flew under the radar a little bit. Sakana AI has created what they call the AI scientist. This is fully automated, open-ended scientific discovery. That means that this model has been trained, and the framework has been trained to actually make scientific discoveries, not

simply repeat back what's in its training set is actually able to make new discoveries, which, if you remember back to Leopold Ashen Brenner's paper, situational awareness, this is the last step before we have the intelligence explosion, artificial intelligence that can actually self-improve. That is what we're seeing here. So, let's read a little bit about it. Today, we're excited to introduce the AI scientists, the first comprehensive system for fully automatic scientific discovery, enabling Foundation models such as large language models to perform research independently. We propose and run a fully Aid-driven system for automated scientific discovery applied to machine learning research. Now, the fact that it's applied to machine learning research is interesting because that means it can self-improve. It could discover ways to get better, and then that compounding effect will be exponential. It is really interesting and a little bit scary to think about this. This really is the last step before the intelligence explosion. And for those of you not aware, Sana AI is a Cutting Edge AI firm out of Japan. The AI

scientist automates the entire research life cycle, from generating novel research ideas, writing any necessary code, and executing experiments, to summarizing experimental results, visualizing them, and presenting his findings in a full scientific manuscript. We also introduced an automated peer review process to evaluate generated papers, write feedback, and further improve results. It is capable of evaluating generated papers with near-human accuracy. So cool! It is able to iteratively develop ideas in an open-ended fashion and add them to a growing archive of knowledge, thus imitating the human scientific community. So basically, it creates an idea or comes up with a hypothesis, tests it, validates it, has a peer review process, and then when everything looks good, it adds it to its own knowledge base so that it could use that for future research. This is stunning. And here's the interesting part: each idea is implemented and developed into a full paper at a cost of approximately $15 per paper. While there are still occasional flaws in the papers produced by this first version, the cost and promise of the system shows that the

potential of the AI scientist is there, so definitely check this out. Again, another paper I haven't yet read in full. There are just so many things that happened this week, but if you want to see a full review of this again, let me know in the comments. All right, next, a quick story from OpenAI introducing sbench verified. We're releasing a human-validated subset of sbench that more reliably evaluates AI models' ability to solve real-world software issues. So, this is a new benchmark, apparently a higher quality benchmark for software engineering models. Each sample in the sbench test set is created from a resolved GitHub issue in one of 12 open-source Python repositories in GitHub. So, here's some example projects tested against this new sbench: Agent List, Auto Code Rover, Melist Tools, AER, one of my favorites, and Sweet Agents. And you can see across the board how they do, so be sure to check this out on open.com. Next, Google had their Gemini live event. I made an entire breakdown video about it. I'll drop that in the

description below, and aside from a couple of cool things they announced, they actually had a pretty Touch and Go part, which was the Gemini live demo, and it actually failed twice. It gave an error, and he had to try it again, failed again, and then used a different phone. Good job, Google, having a backup phone there, and then it finally worked. So, I'm going to play this video. It's a little bit over a minute and a half long, and it shows this almost awkward part of the event where it really almost just completely broke. So, if I happen to come across this concert poster for Sabrina Carpenter, I'll just open Gemini, take a photo, and ask, Check my calendar and see if I'm free when she's coming to San Francisco this year. Gemini pulls relevant content from the image, connects with my calendar, and gives me the information I'm looking for. Oh, looks like we had a little demo issue. Let me try one more time. All right, check my calendar and see if I'm free when she's coming to San Francisco this.

year. Let's see if the demo Spirits are with us today. Nope. Uhoh. Looks like they're not with us. All right. Let me let me uh let me uh try that one more time on another device. Sorry about this folks. Oops. Let me go back. All right. I'm going to take that photo one more time. Check my calendar and see if I'm free when she's coming to San Francisco this year. All right. There we go. All right. Now here's the interesting thing. Google beat OpenAI to market with a full voice model. You can have a conversation with it. It will talk back with you. You can interrupt and it sounds really good. It doesn't quite sound as good as GPT-40 voice but it still sounds good. And you know what? They actually shipped. They actually brought it to market. So good job Google on that. Next, Anthropic released prompt caching with Claude. Caching is one of the most underutilized and highly valuable pieces of functionality that you can add to a large language model. It reduces your cost, increases speed, increases

consistency. If you're building anything with AI and you're doing so at scale, you need caching. So now anthropic has it. When to use prompt caching. Conversational agents reduce cost and latency for extended conversations, especially those with long instructions or uploaded documents. Coding assistance. Large document processing. Detailed instruction sets. Agentic search and Tool use. I mean, this is the most obvious one. Why would you have an agent call an LLM to pull a tool or write a tool when it already should just have direct access to the cached version? Something we know it works. Talk to books, papers, documentation, podcasts, transcripts, and other long-form content. So really cool, and look at these cost reductions. So chatting with a book, 90% cost reduction. Many-shot prompting, 86% cost reduction. Multi-turn conversation, 53%. So this is tremendous, and as you can see here, also the latency has been reduced greatly. So I'm a huge proponent of caching. Figure out where it makes sense for you. Caching is a hard problem, by the way. It's a hard problem to solve from

the engineering perspective, but it's also a hard problem to know what parts of your prompts and responses you should cash. So, a lot of testing in the future, and I'm very, very bullish on cashing in general. Next, apparently, Apple is building a robot, but not really. It's going to be a screen, and it's going to have a robot arm. It's all rumors right now, so using search GPT, Apple is reportedly working on a new device that combines an iPad-style display with a robotic arm. This device, expected to be launched around 2026 or 2027, aims to serve multiple functions such as smart home control, video conferencing, and home security monitoring. But what does it need the arm for? That is weird, but we'll see. The device will feature a thin robotic arm that allows the display to tilt and rotate 360°. So, basically, the arm looks, at least from rumors, to be used for essentially propping up the screen to make sure it's always facing you, or maybe it needs to look at

something, and so it moves the screen and really the camera to look at whatever it needs to look at. And last, Naous Research has released Hermes 3, a fine-tune set of models based on Llama 3.1. Hermes 3 is available in three sizes: 80, 70, and 405B parameters. It has improvements across the board, with particular capability improvements in roleplaying, generative tasks, and more reliable function calling, multi-turn chats, long context coherence, and more. They have a whole research paper about it. I haven't played around with it yet; this week was just too busy to keep up with everything. So again, another thing that if you want me to check it out, if you want me to do a deep dive, let me know in the comments. Here it is versus its relevant Llama 3.1 model. So here's Hermes 345B versus Llama 3.1 405B, and as you can see, it actually looks pretty even in performance, has some wins at certain benchmarks, but loses in some others, and that looks to

be the case across the board, and it has a focus on aligning the model to you instead of a company or external policy. Less censorship, more steerability. I love that. So, if it's going to have similar performance except no sensors, I'm all for it. So, that's it for the news today. There has been so much going on this week. I'm so excited to see what comes next week. Make sure to check out the stream next week, 10:00 a.m. Pacific on Friday. And we just launched memberships for the channel, so we're going to be adding new perks, but there's already some perks in there. If you want to support the channel, be sure to become a member. It really helps, and I appreciate it, and my team appreciates it, and the team is growing, so definitely check out memberships. Thank you in advance, and if you enjoyed this video, please consider giving a like and subscribe, and I'll see you in the next one.

Massive Leap Toward AGI: AI Scientist, Grok 2, SearchGPT, Agent Q, New Coding Model

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

이번 주에는 새로운 로직 및 추론 모델인 Grock 2 베타에 대한 홍보로 3만 명의 트위터 팔로워를 확보한 스트로베리 씨를 비롯해 AI 관련 활동이 급증했습니다.

엘론 머스크의 AI 회사 X에서 개발한 Grock 2는 고급 채팅, 코딩 및 추론 기능을 제공할 예정이며, Grock 2 Mini는 이미 출시되었습니다.

이 모델의 텍스트-이미지 변환 기능은 오픈 소스인 Flux를 기반으로 합니다.

하나는 무검열과 고품질 출력으로 많은 화제를 불러일으키고 있습니다.

터무니없는 이미지 제작 및 AI 발전

도널드 트럼프가 다스베이더가 되어 화성 탐사선을 운전하는 등 도널드 트럼프가 등장하는 기괴한 이미지 창작물이 많이 있습니다.

텍스트-이미지 변환 도구인 Grok 2는 저작권을 무시한 파격적인 결과물로 주목을 받고 있습니다.

한편, 최신 정보와 미디어를 제공하는 Search GPT의 얼리 액세스는 사용자들이 구글에서 다른 곳으로 이동하고 있으며, 멀티온의 에이전트 Q는 계획 및 자가 치유 기능을 갖춘 새로운 세대의 AI 에이전트를 선보이고 있습니다.

AI 모델의 혁신: 라마 3, 코사인 지니, 사카나 AI

라마 3 모델의 제로 샷 성능은 자율 데이터 수집 후 하루 만에 18%에서 81%로 향상되었습니다.

코사인 지니의 소프트웨어 엔지니어링 모델은 S 벤치에서 30%, U 라이트에서 50%의 최첨단 점수를 획득하여 경쟁사를 크게 앞질렀습니다.

사카나 AI는 완전 자동화된 개방형 과학적 발견을 위해 설계된 모델인 AI 과학자를 도입했습니다.

AI 과학자: 과학적 발견의 판도를 바꾸는 사람

AI Scientist는 전체 연구 수명 주기를 자동화하여 모델이 독립적으로 과학적 발견을 수행할 수 있도록 지원하는 획기적인 시스템입니다.

머신 러닝 연구에 적용되어 자기 개선과 기하급수적인 성장을 가능하게 합니다.

또한 이 시스템에는 자동화된 동료 검토 프로세스가 포함되어 있으며, 개당 15달러의 비용으로 아이디어를 정식 논문으로 발전시킬 수 있어 연구에 혁신을 가져올 수 있는 잠재력을 보여줍니다.

구글의 이벤트 하이라이트와 앤트로픽의 프롬프트 캐싱

Google의 Gemini 라이브 데모는 시작부터 순탄치 않았으며, 백업 휴대폰으로 작동하기 전까지 두 번이나 실패했습니다.

이 데모는 이미지에서 관련 콘텐츠를 가져와 사용자의 캘린더와 연결하는 Gemini의 기능을 보여주기 위한 것이었습니다.

한편, 앤트로픽은 다양한 사용 사례에서 비용을 크게 절감하고 대규모 언어 모델의 속도와 일관성을 높여주는 기능인 클로드의 프롬프트 캐싱을 발표했습니다.

애플의 소문난 로봇과 헤르메스 3 출시

애플은 2026년 또는 2027년경에 출시될 예정인 아이패드 스타일의 디스플레이와 로봇 팔을 결합한 디바이스를 개발 중인 것으로 알려졌습니다.

이 장치는 팔이 화면을 받쳐주고 기울이고 회전하는 등 다양한 기능을 제공하는 것을 목표로 하고 있습니다.

나우스 리서치는 역할 놀이, 생성 작업, 보다 안정적인 함수 호출 기능을 개선하여 검열을 줄이고 조종성을 높인 라마 3.

1 기반의 모델 세트인 헤르메스 3를 출시했습니다.