GoKawiil
Tech News
An innovation professor's essay draws on the 2017 'Attention Is All You Need' paper, which introduced the Transformer architecture behind GPT and the generative-AI boom, to argue that enterprise AI has so far only changed individual work. The piece contends a comparable architectural breakthrough is still needed before AI can transform entire organizations rather than just personal productivity.
A research team shared updates on H-Nets, an architecture introduced in 2025 designed to learn language directly from raw bytes rather than relying on hand-crafted tokenizers. The post is the first in a two-part series describing new results and improvements made to the technology since its initial release.
TypeSafe AI, maker of the Jev model, has closed an $870 million funding round at a $7.5 billion valuation led by Andreessen Horowitz, with Sequoia and existing backer DCVC also participating. Jev launched on Sept. 15 and, per the company, is already used by a third of Fortune 500 firms. Unlike large language models, Jev uses a transformer architecture to output probabilities or 'calibrated decisions' rather than text, which TypeSafe says makes it faster and more token-efficient for automation tasks.
A researcher evaluated Jev, TypeSafe's 'System One' classifier model built by attaching a classification head to a pretrained transformer, to see how well its output probability distributions match known physical answers. The test used questions with mathematically established solutions, such as the Maxwell-Boltzmann velocity distribution for a particle in thermodynamic equilibrium, comparing Jev's predicted probabilities against the known correct distributions. The author notes the entire experiment series cost less than $4.00 to run.
A developer built PSSA, a non-transformer language model coded from scratch in Rust without any ML framework, using a recurrent state-space layer, an episodic memory bank, and weights that update while the model runs. Tested against a matched-parameter standard transformer on identical data, PSSA reached a training loss of 3.98 versus the transformer's 4.43 over 12.7 million tokens, and generated text about twelve times faster on the same CPU. On a held-out 198,939-token slice neither model trained on, PSSA also scored lower loss throughout every checkpoint.
An open-source project splits a 0.5-billion-parameter 1.58-bit BitNet language model across seven ESP32S3 microcontrollers connected via SPI daisy-chain. One master unit handles tokenization, embeddings and final sampling, while six compute nodes each process four transformer blocks using 1.58-bit attention and MLP layers with KV caching in PSRAM.
A visual explainer tool called Transformer Explainer illustrates how Transformer-based neural networks work, using the 124-million-parameter GPT-2 (small) model as its example. It walks through core components like tokenization, embeddings, attention mechanisms, and Transformer blocks to show how these models predict the next word in a sequence.