Artificial intelligence is trained and programmed with many different modalities.
Large-language models use text. Image models use images. World models use video. Sound models use sound. And so on.
Data are tokenized, then vectorized, then parameterized, to train a model.
When models are trained well, from good data, and connected to machines, they can do almost anything.
Today, models are used to design products, develop software, run business operations, operate robotics, and much more.
The singular nature of models contributes to their power.
They're only used one way, prompts, they only output one thing, completions, and they only have one flaw, they make mistakes.
Mistakes are dangerous, and yet they produce good data.
Developing models means reducing mistakes by improving algorithms, accelerating hardware, and acquiring more data.
Improving algorithms means innovating and optimizing AI's full-stacks of software.
Accelerating hardware means buying better chips.
Acquiring more data means free-scraping and buying it.
...
Data is digital material; material that has properties.
Proerties of quality include objectives, correctness, computational-usage requirements, etc.
Properties of quantities include size, symbolic richness, modularity, etc.
...
Evaluations take place during build-time and run-time, with static and telemetric analyses, and with inference.
Build-time static analysis uses tools like the language-server protocol to gather symbolic information, check correctnesses (lint), determine architectures, etc.
Run-time telemetric analysis gathers tests' (unit, integration, end-to-end, smoke, etc.) results, validates logs, reads computer usages (storage, G/CPU, timings), etc.
Inference is the most powerful form of evaluation as it has all its internal knowledge, can use the aforementioned methods as tools, can reason over the results of analysis programs as well as the source itself, can loop intelligently, etc.
...
To make a commodity, Bitcode implements materialization-by-measurement on a trustless architecture.
Materialization means making AssetPacks, which are .patch files of the synthesized knowledge plus all the measurements, which include the calculated raw knowledge volume, $BTD.
$BTD is Bitcode's fungible tradable token that are minted alongside AssetPacks, serving two critical functions: to establish the normalized scalar of all measurements and to maximumize incentives for buying and selling AssetPacks.
Trustlessness is achieved by being open-source, fully auditable, immutable, provable, and on-chain.
Bitcode's ledger stores the proofs of the deployed canon, the activity journal, and all system state.
...
Bitcode's applications, APIs, and contracts allow anyone to make, buy, and sell their knowledge commodities.
The Bitcode Measuring API supports user-configurable measuring without commoditization.
...
Advanced Engineered Software, Inc. is a technology company devoted to serving the world with technical-knowledge commoditization.
Our flagship product, Bitcode, provides us with one propriety advantage (closed-source): our custom internal models, which are trained on the whole Bitcode Depository, and are definitionally unexposed to the public.
We have access to all deposits and never share, read, nor otherwise expose them externally, but use them internally for improving our exclusive models to better develop all Bitcode systems optimally for our users.
As Bitcode invented and maximally incentivizes knowledge-commoditization, the postive feedback-loop of our full-access allows us to build the best commodization models in the world for the network, its products, and its participants.
Advanced Engineered Software, Inc.'s primary research areas are knowledge economics, technical-model training and inference methodologies, and cryptographic security.
...