Walter Karshat develops software and the tools used to understand it. His focus spans systems architecture, substantial software projects, and AI coding harnesses.
The record of how software gets built.
Codess discovers and decodes locally retained records from leading commercial AI coding harnesses Claude Code, Codex, and Cursor. It preserves source evidence while harmonizing supported meanings into CoSchema, a common model for investigation.
Its scope includes harness, user, model, tool, MCP, and agent interactions recorded by those systems: prompts and responses, context changes, permissions, delegation, tool inputs and results, file references, and model settings.
Full raw capture is available explicitly; it is not the default. Decoding and relationship reconstruction depend on what each harness retained. Coverage and mapping diagnostics make gaps visible. The relationship chart shows how CoSchema links recorded entities such as sessions, model turns, and tool calls for investigation.
Explore ↗CoSchema preserves vendor distinctions alongside common entities and relationships. The related Skip research explores intermediate representations for schema interoperability; broader transformation tooling remains a research and implementation direction.
CoSchema ↗Schema research ↗Understanding the claims.
Fear AI examines positive and negative claims about AI, employment, and infrastructure. On employment, the useful question is which tasks change, for whom, and over what period. A tool can shorten a task while increasing review effort, or let a team take on more projects. To assess job creation or displacement, we look at hiring, hours, and headcount over time alongside the task measurements.
For infrastructure, compare historical electricity estimates with projections and ask which assumptions drive the gap. National totals describe the scale of demand. Local grid constraints depend on where facilities connect and when they draw power.
Research & publications ↗| Year | TWh | Basis |
|---|---|---|
| 2014 | 58 | Historical |
| 2023 | 176 | Historical |
| 2028 | 325–580 | Scenario range |
Performant and compatible BLAKE2b
uniblake is a portable BLAKE2b implementation with explicit shared-prefix behavior, controlled interfaces, and independent conformance checks.
Each tail produces one digest for the full message, BLAKE2b(prefix || tail). The two lanes show this one-to-one pairing. For consecutive counters, batch hashing writes one independent result per counter.
ub_state_size and ub_state_align, initialize it with ub_init or ub_init_param, then feed the shared bytes once with ub_update.ub_prefix_check to see whether the largest tail fits in the pending final block. When it fits, call ub_hash_tail with the tail bytes, or use ub_hash_n for consecutive 4- or 8-byte little-endian counters. The batch call writes a digest or selected digest slice per counter at the caller’s output stride.ub_copy. Feed the tail with ub_update and finish with ub_final on the copy. The shared state stays reusable, and thread scheduling remains with the caller.