== In-class to do:
clean up destktop space, calendars,
ZOOM, live transcript (start video recording).
ODU course survey.
basic principle, pdf,
MLA: Capacity & efficiency (esp. KV cache reduction / long-context handling).
github: rotary encoding, multihead latent attention,
R1 nature paper, reasoning through RL.
== pre-class to do:
post video:
calendar email invitation:
homework assignment, data camp,
socrative sign in
update Canvas course materials, update learning objectives. assignments as needed:
Test-run code: skip.
kindle book. using ipad to highlight key points.
== In-class to do:
clean up destktop space, calendars,
ZOOM, live transcript (start video recording).
Socrative sign in,
transformer gpt.ipynb on vertex ai,
decoder only
causal masks,
temperature
student presentation: Google Day practice
breakout rooms: course project.
== pre-class to do:
post video
calendar email invitation:
socrative sign in
update Canvas course materials, update learning objectives. assignments as needed:
kindle book.
== In-class to do:
clean up destktop space, calendars,
ZOOM, live transcript (start video recording).
** Google Day presentation;
Socrative sign in,
student presentation;
breakout rooms: course project poster work;
== pre-class to do:
post video of lect 4 LSTM
calendar email invitation:
homework assignment, data camp,
socrative sign in
update Canvas course materials, update learning objectives. assignments as needed:
Test-run code: skip.
kindle book. using ipad to highlight key points.
== In-class to do:
clean up destktop space, calendars,
ZOOM, live transcript (start video recording).
Socrative sign in,
student presentation; Terry and ??
breakout rooms: course project
== pre-class to do:
post video of lect 4 LSTM
calendar email invitation:
homework assignment, data camp,
paper selection, high quality, primary research paper.
potential project (agentic bioinformatics analysis, agentic lab report?, pretraining of transformer, word embedding)
socrative questions (questions on contents from last lecture): TF on VAE
update Canvas course materials, update learning objectives. assignments as needed:
Test-run code: skip.
kindle book. using ipad to highlight key points.
== In-class to do:
clean up destktop space, calendars,
ZOOM, live transcript (start video recording).
Socrative sign in, skipped
Anton: presentation
Normalized flow
breakout rooms: student Kris;
== pre-class to do:
post video of lect 3 GAN
calendar email invitation:
homework assignment, data camp,
paper selection, high quality, primary research paper.
potential project (agentic bioinformatics analysis, agentic lab report?, pretraining of transformer, word embedding)
socrative questions (questions on contents from last lecture): TF on VAE
update Canvas course materials, update learning objectives. assignments as needed:
Test-run code: skip.
kindle book. using ipad to highlight key points.
== In-class to do:
clean up destktop space, calendars,
ZOOM, live transcript (start video recording).
Socrative sign in, skipped
Segio presentation on DoRA
GAN, principle in pdf, then kindle textbook,
breakout rooms: discuss course projects.
Meeting assets for 202510_CS_795_21992_CS 795 gAI Thu Evening are ready!
Meeting summary
The professor outlined requirements for a generative AI course project focusing on ethical and legal considerations, with students needing to submit proposals within a week. The discussion covered various aspects of autoregression models, tokenization in NLP, and the structure and operation of LSTM cells, including how temperature parameters affect model behavior and creativity. The session concluded with a student presentation on weight decomposed low-rank adaptation methods and their performance compared to full fine-tuning, followed by an announcement about breakout rooms for project discussions.
The professor explained that the course project is the only project in the course and must be related to generative AI, with a focus on ethical and legal considerations. He clarified that the project can involve training discriminators or jailbreaking models, but emphasized that the end result should have a generative aspect. The professor also discussed the project proposal requirements, including team composition, background, data sets, AI approaches, and resource needs, and mentioned that students have one week to submit their proposals.
Hong discussed the autoregression model, highlighting that while ChatGPT is a popular example, there are many other models including LSTM and GRU. He explained the concept of tokenization in natural language processing, noting its importance in splitting text into smaller units for analysis. Hong also described the process of embedding, where tokens are represented by continuous floating-point numbers, often trained in the context of input or output, and how this can lead to interesting and meaningful representations.
Hong explained the structure and operation of LSTM (Long Short-Term Memory) cells, focusing on how they differ from traditional recurrent neural networks. He described how LSTMs use a cell state that maintains memory through weighted matrices shared across all time steps, and detailed the four key operations within each LSTM cell: forget, input, cell state update, and output. Hong noted that while LSTMs were a significant improvement over simple RNNs when introduced 28 years ago, they are now considered less efficient than Transformers due to their fixed weight matrices.
The discussion focused on the implementation and behavior of temperature parameters in language models, particularly LSTM models. Hong explained how temperature affects the stochastic nature of model predictions, with higher temperatures leading to more deterministic outputs and lower temperatures increasing randomness. Hamza and Evan clarified that temperature controls the creativity and randomness of model outputs, with Evan confirming this through research. The group also discussed the limitations of Keras for modifying AI models compared to PyTorch, noting its industrial nature and declining usage.
Hong discussed the evolution and modifications of recurrent neural networks, particularly focusing on the LSTM (Long Short-Term Memory) and its limitations. He explained how a modified version of the recurrent unit, known as the Gated Recurrent Unit (GRU), lacks a cell state and memory, which could potentially make it faster to train but less effective in retaining long-term dependencies. Hong also introduced the concept of quantum machine learning, highlighting a recent development where a classical LSTM was combined with a quantum encoder to create a quantum LSTM. He emphasized the potential for quantum computing to revolutionize machine learning and suggested that future generations might need to learn quantum machine learning, even though it is still in its early stages.
Hong and Sergio discussed the first student presentation of the day, agreeing to take a 5-minute break before resuming at 7:05. Sergio confirmed he was ready to present and successfully shared his screen for the presentation.
Sergio presented on weight decomposed low-rank adaptation, introduced by the NVIDIA group, which builds upon LORA (Low-Rank Adaptation) and DORA (Decomposed Low-Rank Adaptation). He explained that while full fine-tuning adjusts all parameters, parameter-efficient methods like LORA and DORA only modify specific components, aiming to replicate full-tuning results with fewer computations. Sergio detailed how DORA decomposes weights into magnitude and direction, allowing independent updates, which leads to a learning pattern closer to full-tuning compared to LORA. Hong and Evan asked clarifying questions about the decomposition process and the implications of the negative slope in the learning trajectory, which Sergio explained as a lack of correlation between magnitude and direction changes. Terry inquired about training time differences between methods, which Sergio did not fully address in the transcript.
Sergio explained the concept of DORA, a parameter-efficient tuning method that reduces the number of parameters by decomposing the weight matrix, resulting in faster training times compared to full tuning. He highlighted that while DORA introduces some computational cost during tuning, it does not affect model latency during inference. Hong inquired about the meaning of scores in the results, and Sergio clarified that higher scores indicate better performance, though the specific metrics are not clearly defined for generative AI. They also discussed the hyperparameters used, including rank (R), which is a key tuning parameter, and Sergio explained how R is chosen based on the results and dimensions of the weight matrix.
The group discussed the performance of DORA and LoRA models, focusing on their efficiency and accuracy compared to full fine-tuning. Sergio explained that DORA can achieve similar or better accuracy than full fine-tuning with fewer parameters, while LoRA performs better with higher ranks but requires more computational resources. The team also explored the concept of quantized models, where the pre-trained model is compressed to reduce memory demands. Hong clarified questions about the ranking system and parameter usage, and the group discussed the implications of different parameter-efficient tuning methods on model latency. Finally, Hong announced that breakout rooms would be set up for students to discuss potential course projects, with 10 rooms available for participants to join
== pre-class to do:
post video of lec 2 VAE.
calendar email invitation:
homework assignment, data camp,
paper selection, high quality, primary research paper.
potential project (agentic bioinformatics analysis, agentic lab report?, pretraining of transformer, word embedding)
socrative questions (questions on contents from last lecture): TF on VAE
update Canvas course materials, update learning objectives. assignments as needed:
Test-run code: skip.
kindle book. using ipad to highlight key points.
== In-class to do:
clean up destktop space, calendars,
ZOOM, live transcript (start video recording).
Socrative sign in, review VAE
== summary, review VAE
GAN, principle in pdf, then kindle textbook,
breakout rooms,
Meeting summary
The meeting began with a review session on variational autoencoders, where students demonstrated good understanding of key concepts including the variational loss function and reparameterization trick. The discussion then moved to Generative Adversarial Networks (GANs), covering their fundamental components, mathematical framework, and training processes, including the challenges and advancements in model training. The latter part of the meeting focused on practical aspects, including the implementation of GANs for image generation, the use of Google Cloud Platform resources like Vertex AI for machine learning applications, and guidelines for course presentations and storage of work.
Hong led a review session on variational autoencoders, confirming that the encoder maps input data to a single latent vector with randomness introduced through auxiliary parameters. Students demonstrated good understanding of concepts like the variational loss function, which includes both reconstruction loss and a regularization term (KL divergence), and the reparameterization trick that allows backpropagation through sampling steps. Hong noted that while some students hadn't signed in, there were 9 confirmed participants, and mentioned that AI meeting note-taking tools were being used by many attendees. The session concluded with a brief mention of moving on to Generative Adversarial Networks in the next lecture.
Hong explained the concept of Generative Adversarial Networks (GANs), which involve a discriminator and a generator. The discriminator aims to distinguish between real and fake data, while the generator creates synthetic data to fool the discriminator. The goal is to reach an equilibrium where the discriminator cannot reliably identify fake data, achieving a 50-50 chance of correct classification. Hong also described the mathematical framework of the value function that guides the training process, highlighting the adversarial nature of the optimization procedure.
Hong explained the mathematical foundation of a binary classifier using cross entropy loss, describing how the value function can be expressed in terms of Kullback-Leibler divergence and Jensen-Shannon divergence between real data and generated distributions. He outlined the training process as a two-step procedure: first maximizing the discriminator using the full loss function, and then minimizing the generator using a simplified version of the loss.
Hong discussed the challenges and advancements in training generative models, focusing on the WGAN with gradient penalty as the current state-of-the-art method. He explained the technical details of the WGAN, including its use of the Earth mover's distance and the introduction of the epsilon parameter for balancing real and fake data. Hong also highlighted the practical implementation of the WGAN using a real-world example involving the detection of fake bricks, which was demonstrated using a dataset of Lego bricks.
Hong explained the structure of a discriminator and generator model for image generation, noting that the discriminator is a convolutional neural network with a sigmoid output for binary classification, while the generator is similar to a variational autoencoder. Hong outlined the training process, which involves computing binary cross-entropy loss for both the discriminator and generator, and mentioned that the optimizer is specified elsewhere in the code. The discussion touched on the technical details of image expansion methods and the inclusion of noise in the loss function to improve model performance.
Hong discussed the implementation and effectiveness of a generative adversarial network (GAN) with a gradient penalty (GP) for image generation. They explained how the GP is calculated and its role in improving the quality of generated images compared to traditional GANs. Hong also introduced the concept of conditional GANs, which concatenate label information to the input and showed that this simple modification can significantly enhance performance.
Hong discussed the evolution of generative AI methods, noting that while the generative adversarial network (GAN) approach was a significant milestone in 2014, the field has since shifted with the advent of agentic AI, which allows for more specialized and sophisticated critiques. Hong also addressed the use of Google Cloud Platform (GCP) and Vertex AI for students in the class, explaining that while GCP provides a range of industrial-level AI tools, the Vertex AI environment is still in its early stages and may require further development. Evan pointed out that the current GCP course focuses mainly on knowledge checks rather than practical use, and Hamza inquired about the speed and capabilities of Vertex AI compared to ODU's supercomputers, to which Hong clarified that the platforms serve different purposes and are not directly comparable.
Terry demonstrated how to access and use Vertex AI, a Google Cloud service for machine learning and AI applications. He explained the difference between on-premises clusters and cloud resources, emphasizing that Vertex AI provides a managed service for model development, training, and deployment. Terry showed the class how to log into Google Cloud using their ODU student accounts and navigate the Vertex AI interface, highlighting key features like the model garden, Vertex AI studio, notebooks, and deployment options.
The meeting focused on discussing the use of Google Cloud Platform (GCP) resources for the course, particularly Vertex AI and storage solutions. Terry explained that a shared project exists for the class, but students should be cautious about deleting each other's work. He demonstrated how to use buckets for storage and recommended copying important data to Git if needed. The group discussed potential future changes to permissions and the possibility of creating individual projects for each student. Hong clarified that presentations should be individual, not group projects, and explained the format and content expectations for presentations. The class was reminded to save their work before the semester ends, as resources may be deleted afterward.
== pre-class to do:
post video of lec 1. done
calendar email invitation: done
homework assignment, data camp,
paper selection:
potential project (agentic bioinformatics analysis, agentic lab report?, pretraining of transformer, word embedding)
socrative questions (questions on contents from last lecture ):
update Canvas course materials, update learning objectives. assignments as needed:
Test-run code: skip.
kindle book. using ipad to highlight key points.
== In-class to do:
clean up destktop space, calendars,
ZOOM, live transcript (start video recording).
Socrative sign in
=> go over assignments, video, datacamp
=> kingma and Weling, 2013 arxiv
=> hqin's proof work
=> further reading, kingma 2019 tutorial
=> play student videos, setup random breakout rooms to discuss presentation papers
== pre-class to do:
calendar email invitation:
syllabus update
socrative questions (questions on contents from last lecture ):
update Canvas course materials, update learning objectives. assignments as needed:
Test-run code: skip.
kindle book. using ipad to highlight key points.
== In-class to do:
clean up destktop space, calendars,
ZOOM, live transcript (start video recording).
Socrative sign in
== summary, went over ch1, touched ch2.
deep generative learning and generative AI. Here are some notable sources:
Stanford University: The course "CS236: Deep Generative Models" by Prof. Stefano Ermon provides detailed lecture videos and slides focused on the foundations, challenges, and applications of generative models in image, text, video, medicine, robotics, and more. The course website with slides is available at https://deepgenerativemodels.github.io/ and videos are on YouTube.youtube
Cornell University: The course "CS 6785: Deep Generative Models," taught by Prof. Vadir Kuleshov, offers an introduction to deep generative models, recent advances, algorithms, and applications including NLP and biology. The lectures are recorded and available on YouTube.youtube
MIT (Massachusetts Institute of Technology): The "Introduction to Deep Learning 6.S191" program covers deep learning basics along with generative AI applications in media, vision, NLP, and biology. All lecture slides, labs, and code are open-sourced and free to use, accessible at https://introtodeeplearning.com/ with lecture videos like "Generative AI for Media" by Google’s Doug Eck on YouTube.introtodeeplearningyoutube
Harvard University: There are presentations specifically on generative AI’s role in education, including outlines for PowerPoint slides targeting its teaching and learning impact, available in PDF form (e.g., from Harvard AI Sandbox materials).hcsra.sph.harvard
University of Virginia (UVA): UVA SEAS offers collections of slides on the technical foundations of generative AI with practical uses in engineering design and analysis.teaching.virginia
Other resources:
NVIDIA Deep Learning Institute has a teaching kit for generative AI with lecture slides, labs, and Jupyter notebooks focused on GPU-accelerated generative AI development.developer.nvidia
Stony Brook University provides teaching resources on generative AI including PowerPoint slides for educators.stonybrook
If you want ready-to-use lecture slides or full course materials, the Stanford CS236, Cornell CS6785, and MIT 6.S191 courses are among the most comprehensive and authoritative sources from major universities. Their materials are typically publicly available online for educational use.
Would you like direct access links, or specific slide decks on any of these?
Beyond Attention: Toward Machines with Intrinsic Higher Mental States
https://techxplore.com/news/2025-05-architecture-emulates-higher-human-mental.html#google_vignette
Meng, K., Bau, D., Andonian, A. & Belinkov, Y. Locating and editing factual associations in GPT.
Adv. Neural Inf. Process. Syst. (2022). at
<https://proceedings.neurips.cc/paper_files/paper/2022/hash/6f1d43d5a82a37e89b0665b33bf3a182-
Abstract-Conference.html>
26. Meng, K., Sharma, A. S., Andonian, A., Belinkov, Y. & Bau, D. Mass-editing memory in a
transformer. in International Conference on Learning Representations (arxiv.org, 2023). at
<https://arxiv.org/abs/2210.07229>
27. Mitchell, E., Lin, C., Bosselut, A., Manning, C. D. & Finn, C. Memory-Based Model Editing at
Scale. in Proceedings of the 39th International Conference on Machine Learning (eds. Chaudhuri,
K., Jegelka, S., Song, L., Szepesvari, C., Niu, G. & Sabato, S.) 162, 15817–15831 (PMLR, 17--23
Jul 2022).
28. Hartvigsen, T., Sankaranarayanan, S., Palangi, H., Kim, Y. & Ghassemi, M. Aging with GRACE:
Lifelong Model Editing with Discrete Key-Value Adaptors. in Advances in Neural Information
Processing Systems (2023). at <https://arxiv.org/abs/2211.11031>
29. Mitchell, E., Lin, C., Bosselut, A., Finn, C. & Manning, C. Fast model editing at scale. in
International Conference on Learning Representations (arxiv.org, 2022). at
<https://arxiv.org/abs/2110.11309>
30. Sinitsin, A., Plokhotnyuk, V., Pyrkin, D., Popov, S. & Babenko, A. Editable Neural Networks. in
International Conference on Learning Representations (2020). at <http://arxiv.org/abs/2004.00345>
31. De Cao, N., Aziz, W. & Titov, I. Editing Factual Knowledge in Language Models. in Proceedings of
the 2021 Conference on Empirical Methods in Natural Language Processing 6491–6506
(Association for Computational Linguistics, 2021).
32. Zhong, Z., Wu, Z., Manning, C. D., Potts, C. & Chen, D. MQuAKE: Assessing Knowledge Editing
in Language Models via Multi-Hop Questions. arXiv [cs.CL] (2023). at
<http://arxiv.org/abs/2305.14795>
33. Cohen, R., Biran, E., Yoran, O., Globerson, A. & Geva, M. Evaluating the ripple effects of
knowledge editing in language models. Trans. Assoc. Comput. Linguist. 12, 283–298 (2023).
De Cao, N., Aziz, W. & Titov, I. Editing Factual Knowledge in Language Models. arXiv [cs.CL]
(2021). at <http://arxiv.org/abs/2104.08164>
35. Meng, K., Sharma, A. S., Andonian, A., Belinkov, Y. & Bau, D. Mass-Editing Memory in a
Transformer. arXiv [cs.CL] (2022). at <http://arxiv.org/abs/2210.07229>
36. Mitchell, E., Lin, C., Bosselut, A., Finn, C. & Manning, C. D. Fast Model Editing at Scale. arXiv
[cs.LG] (2021). at <http://arxiv.org/abs/2110.11309>
37. Hartvigsen, T., Sankaranarayanan, S., Palangi, H., Kim, Y. & Ghassemi, M. Aging with GRACE:
Lifelong Model Editing with Key-Value Adaptors. (2022). at
<https://openreview.net/pdf?id=ngCT1EelZk>
Language Models: Problems, Methods, and Opportunities. arXiv [cs.CL] (2023). at
<http://arxiv.org/abs/2305.13172>
41. Hase, P., Hofweber, T., Zhou, X., Stengel-Eskin, E. & Bansal, M. Fundamental problems with model
editing: How should rational belief revision work in LLMs? arXiv [cs.CL] (2024). at
<https://scholar.google.com/citations?view_op=view_citation&hl=en&citation_for_view=FO90FgM
AAAAJ:M3ejUd6NZC8C>
42. Cheng, S., Tian, B., Liu, Q., Chen, X., Wang, Y., Chen, H. & Zhang, N. Can We Edit Multimodal
Large Language Models? in Proceedings of the 2023 Conference on Empirical Methods in Natural
Language Processing (eds. Bouamor, H., Pino, J. & Bali, K.) 13877–13888 (Association for
Computational Linguistics, 2023).
Here are the URLs for the specified papers:
1. **Locating and editing factual associations in GPT**
Meng, K., Bau, D., Andonian, A. & Belinkov, Y. (2022).
[Link to Paper](https://proceedings.neurips.cc/paper_files/paper/2022/hash/6f1d43d5a82a37e89b0665b33bf3a182-Abstract-Conference.html)
2. **Mass-editing memory in a transformer**
Meng, K., Sharma, A. S., Andonian, A., Belinkov, Y. & Bau, D. (2023).
[Link to Paper](https://arxiv.org/abs/2210.07229)
3. **Memory-Based Model Editing at Scale**
Mitchell, E., Lin, C., Bosselut, A., Manning, C. D. & Finn, C. (2022).
[Link to Paper](https://proceedings.mlr.press/v162/mitchell22a.html)
4. **Aging with GRACE: Lifelong Model Editing with Discrete Key-Value Adaptors**
Hartvigsen, T., Sankaranarayanan, S., Palangi, H., Kim, Y. & Ghassemi, M. (2023).
[Link to Paper](https://arxiv.org/abs/2211.11031)
5. **Fast model editing at scale**
Mitchell, E., Lin, C., Bosselut, A., Finn, C. & Manning, C. D. (2022).
[Link to Paper](https://arxiv.org/abs/2110.11309)
6. **Editable Neural Networks**
Sinitsin, A., Plokhotnyuk, V., Pyrkin, D., Popov, S. & Babenko, A. (2020).
[Link to Paper](http://arxiv.org/abs/2004.00345)
7. **Editing Factual Knowledge in Language Models**
De Cao, N., Aziz, W. & Titov, I. (2021).
[Link to Paper](http://arxiv.org/abs/2104.08164)
8. **MQuAKE: Assessing Knowledge Editing in Language Models via Multi-Hop Questions**
Zhong, Z., Wu, Z., Manning, C. D., Potts, C. & Chen, D. (2023).
[Link to Paper](http://arxiv.org/abs/2305.14795)
9. **Evaluating the ripple effects of knowledge editing in language models**
Cohen, R., Biran, E., Yoran, O., Globerson, A. & Geva, M. (2023).
[Link to Paper](https://transacl.org/ojs/index.php/tacl/article/view/3736)
10. **Language Models: Problems, Methods, and Opportunities**
(2023).
[Link to Paper](http://arxiv.org/abs/2305.13172)
11. **Fundamental problems with model editing: How should rational belief revision work in LLMs?**
Hase, P., Hofweber, T., Zhou, X., Stengel-Eskin, E. & Bansal, M. (2024).
[Link to Paper](https://scholar.google.com/citations?view_op=view_citation&hl=en&citation_for_view=FO90FgMAAAAAJ:M3ejUd6NZC8C)
12. **Can We Edit Multimodal Large Language Models?**
Cheng, S., Tian, B., Liu, Q., Chen, X., Wang, Y., Chen, H. & Zhang, N. (2023).
[Link to Paper](https://arxiv.org/abs/2305.14795)
Citations:
[1] https://proceedings.neurips.cc/paper_files/paper/2022/hash/6f1d43d5a82a37e89b
Quantum Vision Transformers
quantum implementation of attention mechanism
Recent work on representing “Feynman diagrams as computational graphs” has sparked an intriguing idea: Let’s map AI computation to Feynman diagrams to visualize and optimize AI architectures.
💡 By leveraging Meta’s LLM Compiler, we can create a powerful interpreter between quantum field theory techniques and AI model design.
𝐇𝐞𝐫𝐞'𝐬 𝐡𝐨𝐰 𝐢𝐭 𝐰𝐨𝐫𝐤𝐬:
1. Represent AI models as Feynman-like diagrams, with nodes as computation units (e.g., transformer blocks) and edges showing data flow.
2. Use the LLM Compiler to analyze these diagrams, suggesting optimizations based on both structure and underlying computations.
3. Instead of integrating traditional LLVMs we swap it out for Meta’s LLM compiler for a multi-level optimization approach:
- 𝐇𝐢𝐠𝐡-𝐥𝐞𝐯𝐞𝐥: LLM-driven architectural changes
- 𝐌𝐢𝐝-𝐥𝐞𝐯𝐞𝐥: Standard compiler optimizations
- 𝐋𝐨𝐰-𝐥𝐞𝐯𝐞𝐥: Hardware-specific tweaks
𝐓𝐡𝐢𝐬 𝐚𝐩𝐩𝐫𝐨𝐚𝐜𝐡 𝐨𝐟𝐟𝐞𝐫𝐬 𝐬𝐞𝐯𝐞𝐫𝐚𝐥 𝐤𝐞𝐲 𝐚𝐝𝐯𝐚𝐧𝐭𝐚𝐠𝐞𝐬:
1. 𝐄𝐧𝐡𝐚𝐧𝐜𝐞𝐝 𝐢𝐧𝐭𝐞𝐫𝐩𝐫𝐞𝐭𝐚𝐛𝐢𝐥𝐢𝐭𝐲: Feynman diagrams provide a visual language for complex AI systems, crucial for debugging and regulatory compliance.
2. 𝐂𝐫𝐨𝐬𝐬-𝐝𝐨𝐦𝐚𝐢𝐧 𝐢𝐧𝐬𝐢𝐠𝐡𝐭𝐬: The LLM's capabilities to compile and optimize models inspired by QFT principles.
3. 𝐇𝐚𝐫𝐝𝐰𝐚𝐫𝐞-𝐚𝐰𝐚𝐫𝐞 𝐝𝐞𝐬𝐢𝐠𝐧: Optimizations can be tailored to specific GPU or TPU architectures, improving efficiency.
4. 𝐈𝐭𝐞𝐫𝐚𝐭𝐢𝐯𝐞 𝐫𝐞𝐟𝐢𝐧𝐞𝐦𝐞𝐧𝐭: Continuous learning from optimization patterns leads to increasingly sophisticated improvements over time.
Of course, there are challenges. Representing very deep networks or handling the complexity of recurrent connections could be tricky. But I believe the potential benefits outweigh these hurdles.
💡 Now, here's where we can take it to the next level: Combine this Feynman diagram approach with LLM-based intelligent optimization, like Meta's LLM Compiler. We could create a powerful system where both human designers and AI systems work with the same visual language.
🪄 Imagine an LLM analyzing these AI Feynman diagrams, suggesting optimizations, and even generating or modifying code directly. This could bridge the gap between high-level model architecture and low-level implementation details, potentially leading to more efficient and interpretable AI systems.
This approach could be particularly powerful in domains like hashtag#explainableAI and hashtag#AIsafety, where understanding the decision-making process is crucial.
I'm incredibly excited about this direction. It could be a major leap towards more intuitive and powerful ways of developing AI, bringing together experts from physics, AI, and visual design.