There is still a great deal of work left to be done in ML. If you cannot see that, you have a limited imagination.
13
I haven’t left my apartment in weeks. All I do is work on this ML project, from the moment I wake up until I go to sleep.
This is the same pattern of behaviour that caused my wife to leave previously, and I cannot really blame her.
It is also very cold outside.
At this point, I feel compelled to conserve whatever energy I have for this work and whatever money I can spare for food to feed us. We ran out of almost everything last night.
All we have left are three eggs, one piece of toast, about 50 grams of pyttipannu (which my wife hates), one tortilla, and some old porridge from the cupboard.
I can go several days without eating when I am deeply absorbed in work. I can get by on cans of tuna for days.
Of course, it wasn’t always this way. These are the leanest times we have experienced in many moons.
I can tell my wife is reaching her limits.
Lately, I have been working for twenty hours straight or longer.
It really isn’t that difficult for me.
The risk is that, even after completing the project, there may still be no profit or employment to show for it.
It makes me wonder whether the work itself, and the challenges inherent in these circumstances, are the rewards I am ultimately seeking.
I’ve become a gambling man just like my father, grandfather, and great-grandfather. We’ve just gambled in our own ways. I was never interested in Las Vegas casinos, as they seemed to be. My interest was always in technology, and technology alone. Even from within music.
20
The world’s leaders all look like reptilians in human disguises. Not suggesting they are, only observing the resemblance.
23
GAVRIEL.TECH retweeted
Lines leave the positive charges and end on the negative ones, and a bigger charge sends out more of them.
The little crosses mark spots where the field cancels to zero. Watch them slide around as the charges move.
Two panels add a swirl at each charge, so the lines spiral in. The last one adds equipotentials, and they cross every field line at a right angle.
Eight pictures, one formula behind all of them.
NEUROMORPH NANO
Status, 2026-11-17
SCOPE
A digital organism with its own
anatomy: a sensory system that
reads any material (code,
documents, data), a Globalized
Context Stack that coordinates
and selects, and a Core Brain of
MicroModels that processes.
Three systems work together: the
context system, the memory
system, and continuous learning
held in cartridges.
First application: expanding and
compressing Claude Code's context
over MCP, so it understands more
of a project while reading less.
It also serves the user's own
models and agents.
IMPLEMENTED
- Sensory path: drones and eight
coordinated tentacles with
non-overlapping windows.
- Globalized Context Stack: g0 per
tentacle, G1 merge and map, G2
selection and orders, G3.
- Motor path: ledgered orders,
sensory and behavioural data on
separate channels.
- Core Brain: four MicroModels
(reranker, reader, embedder,
writer) on an in-house Rust
engine with CUDA kernels,
verified bit-exact against
recorded references.
- Latency: cold first request in
seconds, warm requests
sub-second.
- Memory system: tiered retention
by run-days, sources re-read on
recall, outcomes recorded from
Claude Code's actions.
- Integration: MCP server and
hooks for the CLI and desktop
app; tools invoked unprompted.
- Base cartridges for all four
MicroModels: seated at start,
open-weights fallback, stored by
kind and model, versioned, with
rollback from off-site backup.
ACTIVE WORK
- Selection precision. Intake
recall is high; candidate
selection is the bottleneck.
Failures traced to origin and
fixed upstream first: request
parsing, scoring calibration,
ranking order.
- Semantic retrieval on every
request, closing the gap between
plain-language requests and code
identifiers.
- Writer base cartridge
verification; batched, cached
reranker on close calls.
NEXT
- End-to-end evaluation in Claude
Code, with and without the Nano,
after this change set.
- Re-map each base cartridge
against its role in the flow.
- Learning layers, then adapter,
learned and preference
cartridges per the design.
1
55
My Finnish social benefits as an unemployed American citizen with legal residency in Finland. I have contributed more in American capital to the Finnish economy than I have ever received in return.
The circumstances are dire, and the only viable option at this point is to build a product and bring it to market as quickly as possible.
I find it increasingly difficult to see any reason to ever pay taxes to anyone, ever.
47
Perhaps I am just particularly unlucky, but Finland seems to work well primarily for those who are mentally ill or seeking refuge or asylum.
This may help explain Finland's high rates of mental illness. Those who are healthy and capable of working are subjected to punishing bureaucracy and taxation, while those who become increasingly mentally ill are rewarded with early, paid retirement. This is fundamentally incompatible with the way I was raised in the United States.
When combined with unemployment rates of around 17% and the constant micromanagement imposed by the bureaucratic system, the experience becomes psychologically detrimental.
I would rather just live in Finland part-time while working from the United States. With the right arrangements and sufficient capital, that could produce a win-win situation. Otherwise, the situation becomes suffocating.
Social healthcare is also overrated. I have yet to receive any care and would rather pay out of pocket for quality service.
The things I do love about Finland are its people, culture, and education system.
1
125














































