Investigations
MLX beats GGUF on a Mac, just
Gemma 4 31B: 12.6 tokens a second against 11.7 on our M4 Pro.
7 minResearch
Investigations
LM Studio’s MLX engine reads prompts twice as fast
Qwen3.6 35B-A3B took in 20,000 tokens in 30 seconds on MLX and 75 on llama.cpp.
7 min
Insights
What Anthropic’s usage data actually holds
We opened the 86 files behind the Economic Index. One row is a metric value, never a chat.
7 min
Explained
Why does a model forget what you told it?
Continual learning, and why no deployed model has it yet.
7 min
Insights
No model card states the compute threshold
The EU AI Act draws its line at 10^25 operations, California and New York at 10^26. Three open-weights cards, none says.
7 min
Investigations
ArtCraft phones home before you sign in
Google Tag Manager and Microsoft Clarity load first. The source has no switch to turn them off.
7 min
Insights
What is the World Humanoid Robot Games?
A yearly games in Beijing with referees, medals and a 100 metres final; the next one is August 2027.
6 min
Insights
Two companies claim first at the robot games
Three medal tables, three winners. We read the result sheets.
8 min
Insights
Five robots beat Usain Bolt in one race
World Humanoid Robot Games 2026 from the result sheets: 666 teams, 2,056 robots, a 100 metres won in 8.64 seconds.
8 min
Insights
Autonomous robots won most of the job events
Hotels, libraries and factories at the World Humanoid Robot Games: by our count, autonomous robots took 13 of the 21 finals.
8 min
Insights
Were the robot games remote controlled?
The rules, event by event: where a human could take the controls, the half-points rule, and four results cancelled.
7 min
Investigations
Humanoid robots broke three human records, not four
Each 2026 Games record claim checked against the rules: three hold on the clock, the high jump had no bar.
9 min
Investigations
Claude Code on a local model mostly describes its tools
Pointed at local hardware it works; 74 percent of the 29,424 tokens it sends before starting are tool definitions.
7 min
Insights
What the 1X NEO home robot sends, in 1X’s words
The FAQ says what leaves the house and when a 1X Expert can see in. The privacy policy never mentions the robot.
6 min
Insights
You can buy a humanoid robot now, from two makers
Unitree sells today and 1X takes deposits. Figure, Boston Dynamics and Agility published a rate, a sold-out year and 2027 dates.
6 min
Insights
Three robot makers give seven answers on lifting
Atlas, Digit 5 and 1X NEO spec sheets in 2026, and each defines the load differently.
7 min
Insights
A humanoid robot costs from $4,900
Unitree's R1 from $4,900, G1 from $13,500, H2 at $29,900; 1X NEO at $20,000 or $499 a month to rent, from the list prices.
7 min
Insights
Humanoid makers only count robots in their filings
Unitree's prospectus shows 5,632 shipped over three years. Tesla's filing leaves Optimus capacity blank.
7 min
Reports
Sixteen coding agents, and what each one sends home
The Local Agent Telemetry Report 2026 tests them on one Mac: who hears from each, and the settings that make it quiet. Free PDF.
8 min
Reports
Seven local AI tools, and which ones phone home
The Local AI Telemetry Report 2026 tests them on one Mac: what they send and which off switches work. Free PDF and dataset.
7 min
Insights
Skills add knowledge, MCP adds hands
A skill is files on disk; an MCP server is tools over a protocol. The specs, the context cost, and the servers we run.
8 min
Insights
Speculative decoding and MTP are one trick in two places
A separate draft model or draft layers inside the model. The llama.cpp flags and the acceptance rates we measured on two Qwen models.
7 min
Insights
Hermes gives Gemma 4 three times the context it needs
Its own engine loaded 193,280 tokens per slot on a Mac. The log, the estimator line behind it, and the config that avoids it.
7 min
Insights
Qwen3.6 is faster, Gemma 4 answers
On a 48 GB M4 Pro Qwen ran at 50.8 tokens a second but gave a move in 4 of 12 ARC-AGI-3 runs. Gemma gave 11 of 12.
8 min
Insights
How a 27B plays past its own context window
The winning ARC-AGI-3 harness drops old turns, keeps facts in Python and keeps going. Infinite play by eviction, read from the code.
6 min
Insights
ARC-AGI-3 scores are six months out of date
From 0.51 percent at launch to 62.7 on the shared harness and 99.9 on the vendor's, in date order with the set and harness on each.
7 min
Insights
ARC-AGI-2 and ARC-AGI-3 numbers do not transfer
Static puzzles scored on accuracy against interactive games scored on action efficiency. Why 24 percent and 0.51 are not on one scale.
7 min
Insights
Five APIs count reasoning tokens five ways
LM Studio, OpenAI, Anthropic, Gemini and OpenRouter all bill them as output and name, place and cap them differently. One client reads all five.
7 min
Insights
The ARC-AGI-3 human baseline is the median, not the best
All 183 level baselines read from disk: 6 to 578 actions, median 60, level by level.
7 min
Insights
ARC-AGI-3 is not saturated
Three numbers circulate as a solved benchmark and only one compares systems. Which set, which harness, and what each cost.
8 min
Explained
How does ARC-AGI-3 score an agent?
RHAE is not accuracy: your moves against a first-time human, squared. The formula, the baseline, and where paper and code differ.
8 min
Investigations
A system prompt can be five times the context window
Four real context files add up to 38,922 tokens: 475 percent of an 8K window and 3 GB of cache.
8 min
Investigations
Two models fit in 48 GB; the speed does not
Both loaded in 26.78 GB of RAM, and generating together dropped each one to about 9 tokens a second.
7 min
Investigations
A 64K context costs 4.4 GB your model manager hides
Measured on a 48 GB Mac: 80 KB of RAM per token, and LM Studio reports 18.85 GB whether you load 8K or 64K.
8 min
Investigations
Your prompt benchmark is timing the cache
An M4 Pro reads prompts at 65 tokens a second; repeat the prompt and the cache reports 13,763, which is 182 times too fast.
8 min
Investigations
An Unsloth quant name describes the recipe, not the file
Two Nemotron builds are byte-identical under different names, and neither holds a single IQ1 or IQ2 tensor.
7 min
Investigations
A 4-bit Nemotron is 12 percent 16-bit
NVIDIA's 22.46 GB Lightning GGUF holds 2.67 GB of BF16: the MTP head, which four publishers ship four ways.
8 min
Investigations
Your 4-bit vision model sees in 16-bit
An mmproj GGUF holds the whole vision encoder; four publishers ship it unquantised, and the one that compresses stops at 8.
7 min
Investigations
The quantisation policy is one line of llama.cpp
One rule predicts all 20 promoted blocks across three publishers' builds, and we had called it a choice.
7 min
Investigations
The speculative decoding head is not in your download
An MTP draft GGUF ships as an extra block or a separate 49-tensor file. We read both; the main file has neither.
7 min
Investigations
Pruned MoE builds really do drop experts, and say so
REAP builds remove a quarter of the experts and declare it: 48 of 64 and 205 of 256 in four tensor tables.
7 min
Investigations
Two builds of one model differ by a whole block
One Qwen3.5 GGUF ships 41 blocks and another 40. The extra block is the MTP head, and only one quantiser kept it.
8 min
Investigations
Hugging Face’s telemetry function is never called
An opt-out variable, a send function nothing uses, and the real trail: a user-agent header naming your Python and torch.
7 min
Investigations
Two local AI tools give your machine a permanent name
A naive search says four of five installed tools carry a machine ID; the real answer, read from the code, is two.
8 min
Investigations
Every local AI app ships a crash reporter, switched off
Three checked: the crashpad process holds no socket and has no upload address set.
7 min
Investigations
OpenCode contacts Sentry before you type
A production DSN is baked into the app and it connects on launch. Two other findings turned out not to be real.
8 min
Investigations
The smallest Kimi K3 is 589 GB
Every shard summed: the smallest complete build is 589.5 GB, the family runs to 769, and two builds disagree by 58.
7 min
Investigations
LM Studio does not track you, but it routes every search
What the app sends, read from the wire.
8 min
Investigations
Ollama’s MLX engine has no 32 GB rule
The requirement is not in the source. Routing is by model format, and the only Darwin memory constant is 512 MiB.
7 min
Investigations
A GGUF filename does not say what is inside
Eight builds of one model audited: a file called UD-Q4_K_XL holds no Q4_K, and none had the silent experts people warn about.
9 min
Insights
NVFP4 and MXFP4: same 4 bits, 36 percent apart
Block size and scale format are the only choices, and NVIDIA reports MXFP4 needs 36 percent more tokens for equal loss.
6 min
Insights
The only NVFP4 numbers favour Q4_K_M
The vendor published nothing to check; the one independent comparison has Q4_K_M ahead on KL divergence with no size win for NVFP4.
9 min
Insights
Does llama.cpp support MLX?
No, and a maintainer said why in 2023. What Apple Silicon runs on, what Ollama did differently, and which to pick.
7 min
Insights
NVFP4: an NVIDIA format that only runs on a Mac
Ollama serves it to macOS alone. Why the 412 error exists, what the measurements show, and why the files stay big.
8 min
Explained
NVFP4 does not range from -8 to 7.5
Its 16 bit patterns hold 15 values, from -6 to 6. Where the wrong number came from and why it matters.
7 min
Investigations
Does your AI phone home?
Seven local AI tools tested, what each one sent captured, and how to check any tool yourself.
8 min
Investigations
How AI tools hide their trackers
A covert tracker in Claude Code, telemetry in others, and how to catch both.
8 min
Investigations
Graphify has 81,000 stars. We read the code instead
Zero telemetry, local-first parsing, an unofficial site, and a package name our tooling refused to install.
8 min
Investigations
Graphify opened 358 sockets, all to LM Studio
We watched every connection over two passes: none left the desk, and the code graph cost zero tokens.
8 min
Investigations
OpenCode or OpenJarvis: which agent phones home?
We captured what each sends. One has a working off switch; the other, the agent we use daily, has none we could find.
7 min
Insights
A brand voice beats a better model
Five files decide how everything your agents write sounds. The output spec, not the model, is the layer that matters.
8 min
Insights
Eight Mac minis can run a 671B model
exo pools Apple Silicon machines into one, so a cluster holds DeepSeek 671B that no single Mac can.
7 min
Insights
Is the M5 Max worth it for local inference?
Memory bandwidth, unified memory ceilings and time to first token against our 48 GB M4 Pro, and whether the upgrade earns its price today.
8 min
Insights
What to run locally and what to send out
A decision guide for 2026: the compliance and cost lines that decide it, with a bank's chatbot as the test case.
5 min
Insights
Gemma 4 E4B makes 8 GB of memory useful
What a small model now does on edge hardware, where it fails, and when it beats reaching for a bigger one.
8 min
Insights
NVFP4 or q4_K_M in Ollama?
The two 4-bit builds are nearly the same size on disk. What the newer format changes for quality and speed, and which to pull.
8 min
Insights
Gemma 4 26B MoE earns its 16 GB
Quality per gigabyte on unified memory, and whether a mixture of experts is worth its real footprint.
8 min
Insights
Gemma 4 31B or Qwen 3.5 27B on a Mac?
Which fits, which is faster, and which answers better when a Mac holds one large model at a time.
9 min
Insights
MLX, llama.cpp or MetalRT on Apple Silicon?
One quantised model, three engines, one Mac: generation speed, memory and setup, measured rather than argued about.
10 min
Insights
LM Studio or Ollama on a Mac?
Setup, speed, model management and the API each exposes, plus which we kept on our own machine and why.
9 min
Insights
What Ollama’s move to MLX changes
Speed, memory and model choice on Apple Silicon, and whether it changes what you should run now.
9 min
Insights
Why we run AI agents offline
Your config files hold keys and endpoints. What a cloud agent can see, and how to keep source code on your machine.
8 min
Explained
What is agentic AI?
One real shift under the word: a model that answers questions became a model that takes actions on your behalf.
10 min
Insights
Local deep research against Perplexity
A research engine on your own machine keeps every query private and charges nothing per report. What it costs you instead.
10 min
Insights
Open WebUI or Jan for a private ChatGPT?
A feature-rich browser workspace or a double-click desktop app. We picked one.
9 min
Insights
Do you need an agent framework?
smolagents and the alternatives, and why most of the noise sells a layer you can skip.
10 min
Insights
The best local model is the one that fits
Not the biggest: the model that leaves headroom in your RAM and matches the job you do every day.
11 min
Insights
Ollama, MLX or Jan on a Mac?
The foundation, the fast one and the friendly one. We run all three for different jobs; here is which fits yours.
11 min
Insights
Aider, Cline, OpenCode or OpenHands?
Four open-source coding agents, four answers to where you want to work: terminal, editor, app window or sandbox.
12 min
Insights
OpenCode, Pi or Cursor?
Three coding agents, three bets on where you work and where your code goes, without the marketing.
12 min
Insights
Why Nemotron beats DeepSeek inside OpenCode
DeepSeek scores higher on coding benchmarks; Nemotron is still the daily driver, and the leaderboard does not say why.
11 min