Scaevolus

Paying 0.055 EUR per GB of traffic sent to French residents makes video streaming fundamentally unprofitable.

Rather than loading the raw approximate decompile code into SQLite, you'll likely get better results using Ghidra directly with tool calls through an MCP adapter. This is especially important when the analysis tracing fails, like with some complex indirect jump patterns with vtables. Xrefs and pattern searches are of course natively supported, and you can load in memory snapshots.

A simple "/goal analyze this codebase, assign names for each function, and creates named structs for parameters." works well, even with quite weak agents. Remember, parsing through heaps of text is their forte!

This setup also works for more exotic non-x86 architectures with more dynamic code loading, like PS2, N64, etc.

"Honesty" is what crossed the threshold for me. It's pretty weird to spend 12 years on something and not bother to write the thousand word history yourself.

Making scalable source control is much easier when you can rely on a distributed database (Spanner) and object storage (GCS/S3/...)!

Reverse engineering and decompilation is not currently blocked by the models, leading to the proliferation of LLM-assisted decomps/recomps. If you ask it to find bugs you're more likely to get a refusal, but simply pointing an LLM at a Ghidra MCP server and telling it to trace program flow, rename functions, or answer questions behaves like normal.

I agree on active devices, and most of the ecosystem is similarly frustrating to deal with.

And why rely on someone else's software for this in 2026? I vibecoded a Rust+Starlark zigbee hub software that uses a few MB of RAM on my router and can directly dispatch HTTP for alerting, as well as exposing prometheus metrics. Zigbee2MQTT apparently wants several hundred MB of RAM.

For passive monitoring, I have six water leak sensors and two temperature/humidity sensors.

The crucial thing is that I can fix it if it breaks. I can debug the low level network packets, support whatever hardware I want, and not have to rely on someone else's cloud ecosystem.

It's great!

Wine-style PS5 emulation has advanced far enough that I would bet money on there being a pirated GTA6 PC release based on the PS5 version available within six months.

Given how women initiate most divorces, it's probably best to set it up so they are the proposing side.

But this isn't raw Gale-Shapley-- the participants aren't ranking all hundreds/thousands of others. In particular, the preference matrix is based on identical mutual compatibility scores, so there's a unique stable solution (not multiple stable ones!), and there's no "side" that benefits, since the solution is identical no matter who does the "proposing".

It creates multiple versions of functions referencing SIMD and lifts the dispatch switching cost to their callers.

> The AST rewrite creates multiple specialized copies of functions, variables, and types that mention simd types, where simd types are replaced with references to size-specialized types in simd/internal/bridge. Each of these bridge types is defined as an archsimd type, but with a restricted set of methods. The specialized functions, variables, and types acquire a suffix of the form @simdNNN, where NNN is either a vector length (128, 256, or 512) or 0, indicating emulation. Functions that mention simd internally, but not in their signature, are converted to wrappers that switch on the SIMD level detected at program start, and call the appropriate specialized version of that function. Specialized functions call other specialized functions directly without dispatch overhead (and perhaps with inlining). This rewrite strategy was chosen as a compromise between code duplication and SIMD performance; the overhead is hoisted as high as necessary to avoid dispatch within SIMD computations, but not higher. If SIMD dispatch appears “too low” in a computation, a gratuitous mention of a simd type will move it upwards, as in this example:

This is a solid improvement over Needle 2, which I tried using for a tool-calling interface to a Runescape database site. Unfortunately it's still not quite capable enough for my target compared to FunctionGemma.

   Model                            Correct tool shape    Exact arguments
  ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━  ━━━━━━━━━━━━━━━━━━━━  ━━━━━━━━━━━━━━━━━
   FunctionGemma fine-tune, BF16       209/230 (90.9%)    196/230 (85.2%)
  ───────────────────────────────  ────────────────────  ─────────────────
   Needle 3 fine-tuned W4A8             74/230 (32.2%)     47/230 (20.4%)
  ───────────────────────────────  ────────────────────  ─────────────────
   Needle 2 fine-tuned W4               59/230 (25.7%)     43/230 (18.7%)

It also compares the total time for jxl-rs's benchmark mode (which does a warmup and a test decode) for 16-bit RGB against 8-bit WebP to 4:2:0 YUV.

Engineers understand that every technology represents inherent tradeoffs.

Did you try blender MCP and prompt for low-poly vertex colored assets?

GPT-6 has a looped transformer / recurrent depth architecture, so it gets some internal CoT reasoning "for free" with no output tokens.

This also makes it much harder to monitor its reasoning.