#language-processing-units

#language-processing-units

Legacy software, often described as 'big balls of mud,' has accumulated over decades, becoming difficult to maintain and understand. These systems rely on extensive teams to function, despite their outdated technology.

Cohere's Transcribe model is designed for tasks like note-taking and speech analysis, supporting 14 languages and optimized for consumer-grade GPUs, making it accessible for self-hosting.

The majority of AI products remain tethered to a single, monolithic UI pattern: the chat box. While conversational interfaces are effective for exploration and managing ambiguity, they frequently become suboptimal when applied to structured professional workflows. To move beyond "bolted-on" chat, product teams must shift from asking where AI can be added to identifying the specific user intent and the interface best suited to deliver it.

By comparing how AI models and humans map these words to numerical percentages, we uncovered significant gaps between humans and large language models. While the models do tend to agree with humans on extremes like 'impossible,' they diverge sharply on hedge words like 'maybe.' For example, a model might use the word 'likely' to represent an 80% probability, while a human reader assumes it means closer to 65%.

A major difference between LLMs and LTMs is the type of data they're able to synthesize and use. LLMs use unstructured data-think text, social media posts, emails, etc. LTMs, on the other hand, can extract information or insights from structured data, which could be contained in tables, for instance. Since many enterprises rely on structured data, often contained in spreadsheets, to run their operations, LTMs could have an immediate use case for many organizations.

Each of these achievements would have been a remarkable breakthrough on its own. Solving them all with a single technique is like discovering a master key that unlocks every door at once. Why now? Three pieces converged: algorithms, computing power, and massive amounts of data. We can even put faces to them, because behind each element is a person who took a gamble.

What happens under the hood? How is the search engine able to take that simple query, look for images in the billions, trillions of images that are available online? How is it able to find this one or similar photos from all that? Usually, there is an embedding model that is doing this work behind the hood.

But tiny 30-person startup Arcee AI disagrees. The company just released a truly and permanently open (Apache license) general-purpose, foundation model called Trinity, and Arcee claims that at 400B parameters, it is among the largest open-source foundation models ever trained and released by a U.S. company. Arcee says Trinity compares to Meta's Llama 4 Maverick 400B, and Z.ai GLM-4.5, a high-performing open-source model from China's Tsinghua University, according to benchmark tests conducted using base models (very little post training).

Semantic ablation is the algorithmic erosion of high-entropy information. Technically, it is not a "bug" but a structural byproduct of greedy decoding and RLHF (reinforcement learning from human feedback). During "refinement," the model gravitates toward the center of the Gaussian distribution, discarding "tail" data - the rare, precise, and complex tokens - to maximize statistical probability. Developers have exacerbated this through aggressive "safety" and "helpfulness" tuning, which deliberately penalizes unconventional linguistic friction.

This process, becoming aware of something not working and then changing what you're doing, is the essence of metacognition, or thinking about thinking. It's your brain monitoring its own thinking, recognizing a problem, and controlling or adjusting your approach. In fact, metacognition is fundamental to human intelligence and, until recently, has been understudied in artificial intelligence systems. My colleagues Charles Courchaine, Hefei Qiu, Joshua Iacoboni, and I are working to change that.

OpenAI has released Open Responses, an open specification to standardize agentic AI workflows and reduce API fragmentation. Supported by partners like Hugging Face and Vercel and local inference providers, the spec introduces unified standards for agentic loops, reasoning visibility, and internal versus external tool execution. It aims to enable developers to easily switch between proprietary models and open-source models without rewriting integration code.