Perplexity CEO Aravind Srinivas, an AI company in the United States, said in an interview with CNBC, "Because DeepSeek must find a way to bypass various restrictions, it actually creates something more efficient in the end. They have proposed many clever solutions." The common perception is that Chinese people are good at copying. But the reality is that some details in DeepSeek 3 are very impressive, and I wouldn't be surprised if Meta borrows some of them and applies them to the Llama model. This is not to say that they are copying, but that they are innovating.

On June 5th, Jinshi Data reported that this morning, Zhipu AI released version 2.0 of the MaaS platform open platform, including a series of new features and models. According to the introduction, the new GLM-4-Flash model masters 26 languages, with a comprehensive ability increase of 40%, surpassing the Llama-3-8B-Instruct in all aspects.

Golden Finance reported that Sam Williams, the founder of Arweave, stated on X platform that AO Super Parallel Computer has made some key breakthroughs in AI technology. Includes: running complete LLMs (such as Llama 3) in AO smart contracts: this means that these AI models can now run directly in smart contracts on the blockchain and make financial decisions; support for WASM64: this allows the protocol layer to use up to about 18 exabytes (EB) of memory, making it easier to process large-scale data; WeaveDrive: this technology allows AO processes to access data on Arweave like a local hard drive. In addition, all details will be showcased at the event on June 20th.

PANews reported on February 28 that, according to The Information, Meta Platforms plans to release LLAMA 3 in July.

Odaily Planet Daily News Curve officially said on the X platform that the one-way lending market based on crvUSD is ready and is waiting for contract audit, and the two-way lending market (collateral assets can be lent) is also basically ready, and the contract writing work has been completed and testing is underway. Previously, it was reported that the 0x7a16 controlled by Michael Egorov, the founder of Curve, has obtained the ENS domain name of llamalend.eth Address, or will launch Llama Lend, an independent lending platform based on crvUSD.

According to a report by the Financial Associated Press on January 19, social media giant Meta CEO Mark Zuckerberg once again reiterated the importance of artificial intelligence (AI). He announced that the company's next goal is to develop AGI (Artificial General Intelligence) and will purchase a large number of NVIDIA AI chips for this purpose. He said Meta is planning to build its own AGI, an AI that can meet or exceed the level of human intelligence in most fields, and pledge to open up the technology to the masses. It was previously reported that Meta is training Llama 3 and will spend billions of dollars to buy Nvidia's artificial intelligence chips.

As reported by CNBC on January 19, Meta is spending billions of dollars on Nvidia's AI chips. Zuckerberg said in an Instagram Reels post on Thursday that the company's AI "future roadmap" calls for it to build a "computing infrastructure at scale." Zuckerberg said the infrastructure will include 350,000 H100 graphics cards from Nvidia by the end of 2024. Zuckerberg didn't say how many GPUs the company had purchased, but the H100 wasn't available until late 2022 and was in limited supply. Analysts at Raymond James estimate that Nvidia's H100 sells for between $25,000 and $30,000, while it could cost more than $40,000 on eBay. If Meta had paid at the low end of the price, the payout would have been close to $9 billion. In addition, Zuckerberg said Meta plans to "responsibly" open source its untapped "general intelligence," which the company is doing for its Llama family of large language models. Zuckerberg said Meta is currently training Llama 3 and working more closely with its Basic Artificial Intelligence Research Team (FAIR) and GenAI Research Team.

As reported by CNBC on January 19, Meta is spending billions of dollars on Nvidia's AI chips. Zuckerberg said in an Instagram Reels post on Thursday that the company's AI "future roadmap" calls for it to build a "computing infrastructure at scale." Zuckerberg said the infrastructure will include 350,000 H100 graphics cards from Nvidia by the end of 2024. Zuckerberg didn't say how many GPUs the company had purchased, but the H100 wasn't available until late 2022 and was in limited supply. Analysts at Raymond James estimate that Nvidia's H100 sells for between $25,000 and $30,000, while it could cost more than $40,000 on eBay. If Meta had paid at the low end of the price, the payout would have been close to $9 billion. In addition, Zuckerberg said Meta plans to "responsibly" open source its untapped "general intelligence," which the company is doing for its Llama family of large language models. Zuckerberg said Meta is currently training Llama 3 and working more closely with its Basic Artificial Intelligence Research Team (FAIR) and GenAI Research Team.

According to qubit reports, Beijing Depth Quest Company launched the first domestic Open Source MoE large model DeepSeek MoE, which is comparable to Llama 2-7B in performance and drops the amount of computation by 60%. The current version of DeepSeek MoE has 16 billion parameters, and the actual number of activated parameters is about 2.8 billion. In addition, the Depth Seek team also revealed that there is a 145 B version of the DeepSeek MoE model in development. Phased preliminary trials have shown that the 145 B DeepSeek MoE has a significant advantage over GShard 137 B, while being able to achieve performance comparable to the intensive version of the DeepSeek 67 B model with 28.5% computational effort.

Odaily Silo Labs has released an update on the proposal to "adjust crvUSD LTV/LT on Silo Llama" on the X platform. The proposal aims to maintain a healthy liquidation margin and only affect users borrowing CRV and YFI. The reduction of crvUSD LTV/LT on Silo Llama is proposed to be executed in the following markets: - CRV-crvUSD: 50%/80% (executed on January 7); - YFI-crvUSD: 55%/70% (executed on January 7); - LINK-crvUSD: 65%/75% (changes implemented), crvUSD lenders can borrow up to 55% of the deposit, and the liquidation threshold is 70%.

1. Meitu's self-developed AI vision model, MiracleVision, will be available to the public through filing 2. Huawei's Noah's Ark Lab and others jointly launched a new large language model architecture, Pangu-π, with the same scale performance exceeding LLaMA 3. Samsung Electronics Announces 2024 Strategic Priorities to Strengthen Chip Business and AI Applications 4. Tsinghua University and the Harvard team launched the AI system LangSplat, which can more accurately describe 3D scenes, 199 times faster than LERF 5. Writerbuddy Releases Global Top AI Tools Report: ChatGPT Creates 14.6 Billion Visits, Ranking No. 1 6. NetEase Youdao Dictionary Announces 2023 Words of the Year:_ 7. IDC: Generative AI spending will double in 2024 to exceed $150 billion by 2027

According to a report by Qubit on January 2, Huawei's Noah's Ark Lab and others jointly launched a new large language model architecture: Pangu-π. By enhancing the nonlinearity, the architecture is improved over the traditional Transformer, which can significantly drop the problem of feature collapse and make the model output more expressive. It is reported that in the case of training with the same data, Pangu-π (7B) surpasses LLaMA 2 in multi-tasking and achieves 10% inference speedup. SOTA up to 1B scale. At the same time, based on this architecture, Huawei has also refined a large financial and legal model called "Yunshan". The work is led by AI Dacheng Tao.

As reported by Xinzhiyuan on January 1, Meta's researcher Martin Signoux made 8 predictions for AI in 2024: 1. Artificial intelligence smart glasses have become a trend. With the rise of multimodal technology, leading AI companies will redouble their efforts to develop AI wearables. What could be better than a glasses form factor to host an AI assistant? 2. ChatGPT is to AI assistants what Google is to search. In 2023, ChatGPT is starting to shine, with Bard, Claude, Llama, Mistral, and thousands of spin-offs coming out one after another. As productization continues to advance, ChatGPT will no longer be the only reference standard in this field, and its valuation will also face revision. 3. Goodbye to large model models, hello multimodal models. LMMs will continue to emerge and replace LLMs in the debate over multimodal evaluation, multimodal safety, multimodal this, multimodal that. In addition, LMM is a stepping stone towards a true general AI assistant. 4. No major breakthroughs, but improvements on all fronts. The new model will not lead to a real breakthrough (GPT-5), LLMs are still limited in nature, and are prone to hallucinations. We won't see any leaps that make them reliable enough in 2024 to "solve the basic AGI." 5. Small models (SLMs) are already emerging, but cost-effectiveness and sustainability considerations will accelerate this trend. Quantitative technology will also be greatly improved, thereby driving a wave of device integration for consumer services. 6. The Open Source model defeated GPT-4, and the battle between Open Source and closed gradually subsided. Looking back at the vitality and progress of the Open Source community over the past 12 months, it's clear that the Open Source model will soon close the performance gap. 7. Benchmarks remain a challenge. Benchmark, leaderboard, or evaluation tools are a one-stop shop for model evaluation. Instead, we will see a series of improvements (such as HELM) and new initiatives (such as GAIA), especially in terms of multimodality. 8. The risks that exist are not much discussed compared to the existing ones. While X-risk makes headlines in 2023, public discussion will focus more on existing risks and controversies related to bias, fake news, user safety, election integrity, and more.

As reported by Reuters on December 12, Meta used pirated books to train its AI model, despite warnings from its own lawyers. According to a new complaint filed Monday, Meta is accused of using a protected work to train an artificial intelligence language model, Llama, without permission. The lawsuit was co-sponsored this summer by writers such as comedian Sarah Silverman, Pulitzer Prize-winning Michael Chabon, and others. This week, the case was joined with another similar lawsuit. Last month, a California judge dismissed parts of Silverman's lawsuit, but said the authors would be allowed to amend the lawsuit. The new complaint, which includes chat transcripts of a Meta researcher discussing access to datasets on a Discord chat server, is evidence that Meta is aware that its use of the book may not be protected by U.S. copyright law. In the chat transcript cited in the indictment, researcher Tim Dettmers describes his argument with Meta's legal department over whether it is "legal" to use book files as training data. Tim Dettmers mentioned that lawyers in Meta's legal department have said that there could be legal issues if the books were used to train AI models. The lawyer said the data could not be used, and if it did, the model could not be published.

According to IT Home on December 8, Meta yesterday launched an AI detection suite called Purple Llama, claiming to provide a series of tools to help developers detect their own AI security and "responsibly deploy generative AI models and applications". The Purple Llama suite currently offers the "CyberSec _" assessment tool, the Llama Guard "Input and Output Security Classifier", and Meta claims that more tools will be added to the suite in the future. Meta introduced that the Purple Llama suite aims to standardize its own Llama language model, and can also act on the AI models of other competitors, and the name of this suite is based on the "purple team concept" in the field of security. Meta claims that the Purple Team is a combination of "attacking the red team and defending the blue team," and that the Purple Llama suite is designed to assess and mitigate potential threats to AI models, "balancing both attack and defense strategies." Specifically, the Purple Llama suite is now capable of quantifying the metrics of cybersecurity risk for large models, assessing possible code vulnerabilities, and providing insights into the security of the model, in addition to inspecting and filtering the input and output of the model to prevent hackers from "jailbreaking" the model in question.

According to the AIGC open community, Microsoft announced the addition of 40 new models to the Azure AI cloud development platform, such as Falcon, Phi, Jais, Code Llama, CLIP, Whisper V3, Stable Diffusion, etc., covering text, images, code, speech, and other content generation. Developers only need to use APIs or SDKs to quickly integrate models into their applications, while supporting tailored functions such as data fine-tuning and instruction optimization. In addition, developers can quickly find the right product in Azure AI's Model Mall by searching by keyword, for example, by typing in the word "code" to display the corresponding model.

According to VentureBeat, on November 21, Microsoft Research released a small language model for Orca 2. It is reported that the performance of Orca 2 in testing complex inference tasks in a zero-RF environment is comparable to that of large language models such as Meta's Llama-2 Chat-70B, or even five to ten times more. Available in sizes of 7 billion and 13 billion parameters, Orca 2 builds on the original 13B Orca model, which demonstrated powerful inference capabilities a few months ago by mimicking the stepwise inference trajectories of larger, more capable models. At present, Microsoft has open-sourced the two new models in order to further investigate how to develop and evaluate smaller models with comparable performance to large models.