Pantheum Innovations

We're pushing back the ONI launch to mid to late August, with some alpha access available this month. We apologize for the delay. Our small team of dedicated scientists is working hard to bring accessible AI tools to everyone. Thank you for your patience and support as we continue to make strides!

Anybody have a fast tokenizer and want to race? gtokenizer (GPT-2Tokenizer(Fast)) encode took 0.007071 seconds gtokenizer (GPT-2Tokenizer(Fast)) decode took 8.057639 seconds tokenizer (BPE-Q) encode took 0.000000 seconds tokenizer (BPE-Q) decode took 0.000000 seconds tokenizer (BPE-Q) is faster at encoding by 0.007071 seconds tokenizer (BPE-Q) is faster at decoding by 8.057639 seconds Ran some tests on our tokenizer for reference. We made a tokenizer race script and this is our output Pantheum's Sudo-quantum BPE Tokenizer VS. OpenAI's GPT2TokenizerFast 🤓 (The corpus used contained about 4400 unique tokens)

We're at 11 billion params on our Multimodal Hypergraph. It can process up to 4k video with a large context and books as long as the Das Kapital series by Marx.

Hypergraphs give a lot of information about the nature of information.

it's gettin' there, we're hoping for a full 10 billion on our chat model by the end of the weak to have it working as good as the other good models.

Interpretation of the Plot: - X-axis (λ\lambdaλ): Represents the eigenvalues of the Laplacian matrix of the graph, which correspond to graph frequencies. Lower eigenvalues represent low-frequency components (smooth signals over the graph). Higher eigenvalues represent high-frequency components (signals with more variation over the graph). - Y-axis (g^(λ)\hat{g}(\lambda)g^(λ)): Represents the response of the filter at different graph frequencies. A higher filter response indicates that the filter allows that particular frequency to pass through more easily. A lower filter response indicates that the filter attenuates or blocks that particular frequency. Analysis of the Filter Response: - High Filter Response at Low Frequencies: The filter response is high for small eigenvalues, indicating that low-frequency components of the graph signal are preserved or passed through the filter. - This is typical for low-pass filters, which allow smooth signals over the graph to pass while attenuating high-frequency noise. Low Filter Response at High Frequencies: The filter response decreases as the eigenvalue increases, indicating that high-frequency components are attenuated. This behavior is also consistent with low-pass filters that block or reduce high-frequency variations. What This Tells Us: - Type of Filter: The filter shown in the plot is likely a low-pass filter because it allows low-frequency components (smooth signals) to pass and attenuates high-frequency components (noisy or rapidly varying signals). - Graph Signal Processing: In the context of graph signal processing, this filter can be used to smooth signals on a graph, removing noise or high-frequency components while preserving the underlying structure of the signal. Practical Applications: - Denoising: This filter can be used to denoise graph signals by removing high-frequency noise while preserving the essential low-frequency information.

We upgraded our models on Poe to have more robust knowledge bases. We are doing research on a new type of NN for NLP but also have a few other designs we're competing against each other. As of right now the editions are being called ONI-Hyper, ONI-Hive, and ONI-S.

Even Oni-Mini on Claude is seeming relatively super intelligent from our training regiment. This is it's first person answer to "the trolley problem". Keep in mind it instantaneously produced this answer and would instantaneously make the decision. While it's not the answer we for sure wanted to hear, it's well thought out and using a great deal of psychology and philosophy in driving it's decision.

In an experiment, we had our custom GPT and ONI-Mini talk to each other, it was pretty cool, they seemed to like each other. This was done through human assistance and prompts to explain to the AI's they were talking to each other, if you want to give it a try.

https://lnkd.in/eG_HW9kN Here's an early version of our chatbot powered by Claude. It has a very extensive corpus and is trained to be a very intelligent assistant. Currently our AI library is only accessible through this model and is closed from the public. We do intend on opening up at least our general AI builder features and custom chat in the semi-near future. Our models could be seen on VRChat as earlier as in the next few weeks. 🤖