Reprint Generate AI
i. GPT-5 inference and RL acceleration come? Altman public GPT-4.5 is ready
1. GPT-4.5 has been implemented in OpenAI. Through inference models and reinforcement learning technology breakthroughs, GPT-5.5 can be achieved without 100 times of computing power;
2. OpenAI plans to launch a fully autonomous intelligent body that can solve all difficult tasks except scientific discovery by the end of the year, supporting multi-modal input and Internet access;
3. Altman predicts that the AI computing power of a single data center will exceed the current total intelligence capacity of the Earth in 2035, but emphasizes that the nature of human beings will not change and the working form will change.
https://mp.weixin.qq.com/s/j9eT84ZgNGgt5GV9oc18kQ
II, DeepMind A new combination of sacrifice neural network and symbol AI AlphaGeometry2
1. DeepMind released AlphaGeometry2 system, which successfully solved 84% of geometric problems in IMO in the past 25 years, surpassing the average level of gold medalists in Mathematical Olympiad;
2. The system combines the hybrid architecture of neural network (Gemini model) and symbol AI to realize more powerful mathematical reasoning ability, while the o1 reasoning model of OpenAI fails to solve any problems;
3. DeepMind breaks through the scarcity of training data by creating 0.3 billion theorems and proven synthetic datasets, but the system still has technical limitations such as being unable to deal with variable pixels.
https://mp.weixin.qq.com/s/9pKbAe6VHunFDLA9nmoW9Q
III. Quilt DeepSeek OpenAI decided to find the limelight in "aesthetics"
1. OpenAI carries out overall brand reshaping, replacing new fonts, new labels and new color matching, and launching exclusive OpenAI Sans fonts and dynamic design of "emotional dots", with the goal of showing a more unified and humanized brand image;
2. The new Logo is optimized into a more regular "kind of life" based on the original "flower". The new font takes circle as the design core, and specially adds irregular design into the perfect circle to embody the humanization concept;
3. This brand reconstruction mainly relies on artificial design rather than AI, and the overall design conveys the idea that "artificial intelligence should enhance rather than replace human creativity.
https://mp.weixin.qq.com/s/-4sk_3eVy14ZQYuB-qOfYw
4. Open-source LLM fine-tuning artifact Unsloth,7GB GPU experience aha moment?
1. Unsloth AI, an open source project, has achieved a major breakthrough. By optimizing GRPO training methods, memory usage has been reduced by 80%, enabling 7GB memory GPU to locally run DeepSeek-R1-level inference models;
2. Unsloth implements deep integration with vLLM, which can increase the model throughput by 20 times while only half VRAM, so that a single 48GB GPU can fine-tune Llama 3.3 70B;
3. The project has won more than 20000 stars in GitHub, and its core team consists of only two brothers, which has successfully lowered the deployment threshold of AI inference models.
https://mp.weixin.qq.com/s/WayXEwbzAv00gd1uj-7jqg
5. OpenAI or Sora image generator under internal testing, code-named "papaya 」
1. OpenAI is testing the image generation function of Sora internally, including video and image generation switching buttons. Video push will be divided into "Best" and "Top" categories;
2. The new image generation function code is "papaya", which may be driven by the existing "sora-turbo" model instead of DALL-E 4;
3. The "Images Internal" category appears on the Sora interface of the Internal test version, suggesting that an independent image generation and display function will be launched soon.
https://mp.weixin.qq.com/s/_VGAJffuQ_V63ZT0GCbUQA
frontier Technology
VI. New breakthrough of Apple robot: Robot's "smart" gesture is as attractive as human beings
1. Apple AIML the research team put forward the framework of EMOTION, combining big language model and visual language model, so that robots can generate natural and situational gestures;
2. EMOTION understands the situation through "context learning" and continuously optimizes gesture performance through human feedback. Experiments show that there is little difference in naturalness and comprehensibility between robot gestures and human gestures;
3. The technology is still facing challenges such as hardware limitations, motion fluency, and computing time. It takes 24-33 seconds to generate an initial sequence of actions.
https://mp.weixin.qq.com/s/eu7GGLHm22-MaXgZKHtVSQ
7. Another breakthrough in AI! Decoding human thoughts, can brain injury patients achieve "barrier-free" communication?
1. Meta released two major brain-computer interface studies. Its AI model can decode up to 80% of brain language signals through non-invasive methods, and the error rate of the best participants is only 19%;
2. The research adopts MEG and EEG technology to record brain activity, and develops Brain2Qwerty deep learning architecture for decoding, which improves performance by 1.14-2.25 times compared with traditional methods;
3. Research reveals that brain language generation follows a hierarchical process, from context to vocabulary, syllables and letters, but this technology still has limitations such as real-time decoding and application scenarios.
https://mp.weixin.qq.com/s/FjfGJ3DM_cF-VZXeYwkJ4w
report views
eight, cloud, Chip, software, car, mobile phone are connected to DeepSeek, what on earth?
1. Nvidia, AMD, Intel and other chip manufacturers have announced the adaptation of DeepSeek. However, due to chip architecture differences, additional software and algorithms need to be developed to achieve adaptation. Nvidia took the lead in launching NIM with CUDA ecological advantages;
2. Mainstream cloud computing service providers at home and abroad have successively provided DeepSeek model deployment tools and API services, expanding market share by lowering the threshold for enterprises to use, and some have also developed value-added services based on DeepSeek;
3. Although terminal manufacturers such as mobile phones and automobiles claim to access DeepSeek, due to the large number of parameters in the original model (671 billion), they can only use small versions with 1.5 billion and 7 billion parameters, the performance is significantly different from that of the original version.
https://mp.weixin.qq.com/s/MnLGHr9n8gU5NxlcvH3_aQ
9. Does DeepSeek have national transport-level innovation? 20000-word hard core analysis V3/R1 architecture
1. DeepSeek has made major breakthroughs through MLA and the improved MoE architecture, increasing the computing efficiency of the model by about 10 times, and optimizing the AI infrastructure close to the upper limit of GPU performance;
2. DeepSeek's R1 model proves for the first time that the reasoning ability can be trained through pure reinforcement learning (R1-Zero) without Manual labeling of data, providing a new possible path for future AI to exceed the upper limit of human thinking;
3. Although DeepSeek's innovation is of great significance, it is not accurate to say that the media excessively hyped it to break CUDA monopoly and lead to a sharp drop in Avida's market value. Its technology is still optimized in Avida's GPU ecosystem, scale-up at the software level.
https://mp.weixin.qq.com/s/0n9lUH9WsDQTGnPCweUkdA
x. DeepSeek's innovation originates from chip blocking, and destropic or become the biggest loser?
1. DeepSeek's success stems from forced innovation: Facing H800 limited computing power, achieved efficiency breakthroughs through architecture innovations such as MLA and MoE, and the training cost was only $5.576 million, which proved that the blockade promoted technological progress on the contrary;
2. The emergence of DeepSeek has a profound impact on the United States: it challenges the U.S. chip blockade strategy, exposes the limitation of "competing by blocking past innovations", and may accelerate the development of AI innovation in China;
3. The impact on the AI industry structure is complex: OpenAI can maintain the consumption-level advantage, but the loss of influpic is the largest; It is beneficial to large technology companies for a long time, because the reduction of AI costs expands the application scenarios.
