iiinova — where your digital self keeps evolving

We put up kelly the skelly today. Here is our family Halloween playlist. open.spotify.com/playlist/1Zc... Happy spooky season, bsky fam.

A 12 ft skeleton in front of oak trees around sunset.

040

Al Merose @al.merose.com · 03/10/2026

Attention within my data model for SQL is slightly verbose, but reads like the mathematical definition. github.com/xqlsystems/d...

010

Al Merose @al.merose.com · 30/09/2026

That different frontier models perform in the lead in agentic coding, depending on which benchmark you look at, means to me that one might want to see which benchmark reflects the work you typically do to choose the model family.

Benchmark results of gemini 4

https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/

030

Al Merose @al.merose.com · 30/09/2026

Getting excited about what I’ve been cooking up recently. These 3 lines of SQL, which look like the mathematical definition, are all you need to define the weight update rule for SGD. This is very close to the way you’d define it in JAX and Python. docs.jax.dev/en/latest/au...

WITH loss AS (SELECT SUM(power (x.v *w.val -y.v, 2)) AS 1
FROM × JOIN w ON x.i = w.i JOIN y ON x.s = y.s)
SELECT w.i, w.val - 0.1 * g.val AS val = g.i
FROM w JOIN grad(loss, w.val) g ON w.i = g.i

040

Al Merose @al.merose.com · 27/09/2026

Having a surprising conversation with Opus 5.5 rn

Claude Opus 5.5 thinking trace: “Working through feelings of regret about...”

060

Al Merose @al.merose.com · 23/09/2026

This is one of my favorite videos on the internet. A classic.

020

Al Merose @al.merose.com · 22/09/2026

Similarly, Zizek’s concerns with AI are not the technology, but the capitalist environment in which it is developed. slavoj.substack.com/p/why-contai...

Part of the threat of AI thus resides not in technology as such but in the social relations within which AI machines are produced: we live in a universe of competition and global suspicion that prevents us from breaking out of the destructive logic of the prisoners’ dilemma. Although we rationally know that solidarity and collaboration are better for all participants, the capitalist focus on profit privileges distrust of others: it perceives others as potential threats who will use our trust as an excuse to attack us even more brutally.

152

Al Merose @al.merose.com · 17/09/2026

Lil’ baby in my lap, lil’ Claude in my hand

Accessing workspace:
/home/alxmrs/git/xarray-sql
Quick safety check: Is this a project you created or one you trust? (Like your own code, a well-known open source project, or work from your team). If not, take a moment to review what's in this folder first.
Claude Code'll be able to read, edit, and execute files here.
Security guide
› No, exit
Yes, I trust this folder
Enter to confirm • Esc to cancel
[main] <tmux] 1:claude* 2:zsh-"nixos" 10:48 • 17-Sep-26

020

Al Merose @al.merose.com · 15/08/2026

Had a dumb little question for claude: what if the mandlebrot style fractal incorporated the fibonacci sequence? The result is this fractal that I hadn’t seen before

140

Al Merose @al.merose.com · 02/08/2026

TL;DR: Here's what Kelp does so far.

100

Al Merose @al.merose.com · 15/07/2026

Claude asked for approval for calling `rm` on a local log file (metadata, meaningless), but did not ask for approval and went ahead and ran `gsutil -m rm` on a data corpus used for training 🙃

I checked - and I have to be straight with you: it's not recoverable from the bucket.

130

Al Merose @al.merose.com · 13/07/2026

I got to rank 7, now oscillating around 12. A dork dream come true.

020

Al Merose @al.merose.com · 13/07/2026

We’re on the front page of HN

16. * Show HN: I implemented a neural network in
SQL (github.com/xqlsystems)
9 points by alxmrs 35 minutes ago | hide | discuss | edit

1120

Al Merose @al.merose.com · 13/07/2026

New twiggy pic just dropped

An unbelievably cute grey tabby cat with her eyes closed resting her head on her paws in a basket like thing.

0240

Al Merose @al.merose.com · 06/07/2026

Here is what the loss looks like:

$% uv run benchmarks/nn.py warning: The Python request from `.python-version` resolved to Python 3.11.11, which is incompatible with the script's Python requirement: `>=3.12` Installed 41 packages in 232ms step 0: loss 2.298 train_acc 0.154 test_acc 0.145 step 5: loss 1.559 train_acc 0.440 test_acc 0.325 step 10: loss 1.150 train_acc 0.642 test_acc 0.610 step 15: loss 0.865 train_acc 0.702 test_acc 0.620 step 20: loss 0.825 train_acc 0.716 test_acc 0.665 step 25: loss 0.686 train_acc 0.752 test_acc 0.750 step 30: loss 0.598 train_acc 0.800 test_acc 0.760 step 35: loss 0.584 train_acc 0.788 test_acc 0.720 step 40: loss 0.484 train_acc 0.864 test_acc 0.815 step 45: loss 0.582 train_acc 0.792 test_acc 0.735 step 50: loss 0.504 train_acc 0.800 test_acc 0.775 step 55: loss 0.452 train_acc 0.830 test_acc 0.730 step 59: loss 0.413 train_acc 0.864 test_acc 0.785 trained (784, 196, 32, 10) MLP; weights -> xarray {'inp_0': 785, 'out_0': 196, 'inp_1': 197, 'out_1': 32, 'inp_2': 33, 'out_2': 10}. <xarray.Dataset> Size: 1MB Dimensions: (inp_0: 785, out_0: 196, inp_1: 197, out_1: 32, inp_2: 33, out_2: 10) Coordinates: * inp_0 (inp_0) int64 6kB 2 3 4 5 6 7 8 9 ... 778 779 780 781 782 783 784 * out_0 (out_0) int64 2kB 9 13 20 27 33 45 48 ... 155 166 168 181 187 189 * inp_1 (inp_1) int64 2kB 0 1 2 3 4 5 6 7 ... 190 191 192 193 194 195 196 * out_1 (out_1) int64 256B 0 2 4 6 8 22 25 28 29 ... 7 16 19 24 3 10 18 21 * inp_2 (inp_2) int64 264B 0 1 2 3 4 5 6 7 8 ... 9 13 20 27 12 14 30 31 32 * out_2 (out_2) int64 80B 1 0 2 4 6 8 3 5 7 9 Data variables: layer_0 (inp_0, out_0) float64 1MB -0.05519 0.05624 ... -0.04738 0.01336 layer_1 (inp_1, out_1) float64 50kB 0.1518 -0.1549 ... 0.02047 -0.05072 layer_2 (inp_2, out_2) float64 3kB -0.04275 0.1687 ... -0.04326 -0.05491$

170

Al Merose @al.merose.com · 01/07/2026

Here is OLS by gradient descent as a single recursive SQL statement

WITH RECURSIVE params (step, a, b) AS (
SELECT 0, 0.0, 0.0
UNION ALL
SELECT params. step + 1,
params.a - 1r*AVG (grad(loss, a)),
params.b - 1r*AVG (grad(loss, b))
FROM params CROSS JOIN d WHERE params.step ‹ STEPS
GROUP BY params.step, params.a, params.b)
SELECT * FROM params ORDER BY step

071

Al Merose @al.merose.com · 28/06/2026

This is where I post from

080

Al Merose @al.merose.com · 28/06/2026

I’m so glad Mythos exists because now I can drink a Mythos and use a $20/mo Claude plan on Opus and rarely hit my token limit.

010

Al Merose @al.merose.com · 27/06/2026

A 🤯 realization I’ve just had on this beach: I think I can create a real general circulation model (a climate model) in SQL. mitgcm.readthedocs.io/en/latest/ex...

120

Al Merose @al.merose.com · 26/06/2026

Second, this is the entire query needed to calculate zonal statistics over a set of regions — here, avg 2m temp over SE Asia, the Sahara, the Amazon, Greenland, and the Australian Outback. This involves joining raster and vector data at once (hard to do with arrays alone).

SELECT r.region, AVG(a."2m_temperature") - 273.15 AS avg_c
FROM era5.surface a JOIN regions r
ON a.latitude BETWEEN r.lat_min AND r.lat_max
AND a.longitude BETWEEN r.lon_min AND r.lon_max
WHERE a.time BETWEEN TIMESTAMP '2020-06-01' AND TIMESTAMP '2020-06-01 23:00:00'
GROUP BY r.region

100

Al Merose @al.merose.com · 26/06/2026

To preview my demo, here are two queries that are now possible with SQL because of Xarray-SQL: First, the core logic of Weatherbench — it calculates the RMSE of two ML weather forecasts (GraphCast and Pangu) as compared to ERA5.

SELECT f.model, f.prediction_timedelta AS lead,
SQRT(AVG(POWER(f."2m_temperature" - e."2m_temperature", 2))) AS rmse
FROM forecasts f
JOIN era5 e
ON e.time = f.time + f.prediction_timedelta -- valid_time = init + lead
AND e.latitude = f.latitude
AND e.longitude = f.longitude
GROUP BY f.model, f.prediction_timedelta

100

Al Merose @al.merose.com · 26/06/2026

Ok, but when will I be able to talk to Twiggy?

Twiggy with her eyes closed on my couch.

240

Al Merose @al.merose.com · 21/06/2026

Finally got to try out Mythos

3920

Al Merose @al.merose.com · 04/06/2026

My favorite thing about my colleague's post on the scaling science behind Marin has to be the part where an open source contributor noticed an opportunity for improvement, quickly made a patch, and that idea made it into a big training run.

Routed Expert Normalization and Rescaling
Marin is open-development: anyone can follow experiments and contribute. Elie Bakouch, a community member, recently noticed a difference between our expert weighting and DeepSeek's*. Within hours of his suggestion to renormalize and scale the routed expert outputs, we confirmed the small boost and added it to the recipe. Without his recommendation, this improvement would not have made it into our next large scale run. When evaluating the change, we compared to the in-progress recipe indicated by the blue dots below.

020

Al Merose @al.merose.com · 21/05/2026

Twiggy, my little art collector

My grey tabby cat sitting over a vibrant painting of an artist’s interpretation of jazz. Painting are on the wall behind her.

030

Al Merose @al.merose.com · 18/05/2026

I just merged in a new update in xarray-sql lead by a UCSD Data Science undergrad. Now, we can use SQL to make queries on ARCO-ERA5, check it out: github.com/alxmrs/xarra...

$import xarray as xr import xarray_sql as xql # Open a year of ARCO-ERA5 — all 273 variables. Selecting a year up front # keeps Dask's partition setup cheap before any chunks are read from GCS. ds = ( xr.open_zarr('gs://gcp-public-data-arco-era5/ar/full_37-1h-0p25deg-chunk-1.zarr-v3', chunks=None, storage_options={'token': 'anon'}) # Anonymous read from the public GCS bucket — no auth required. .sel(time='2020') .chunk({'time': 1}) ) ctx = xql.XarrayContext() ctx.from_dataset('era5', ds, table_names={ ('time', 'latitude', 'longitude'): 'surface', ('time', 'level', 'latitude', 'longitude'): 'atmosphere', }) # Registration: ~0.5s for a full year of hourly ERA5, all variables. # Heads up: ARCO-ERA5 has 262 surface + 11 atmospheric variables. The library # pushes column projection down to Zarr, so SELECT only fetches what you ask # for — but `SELECT * FROM era5.surface` would try to pull every variable # across the year (terabytes from GCS). # ---> Always SELECT specific columns. <--- # Average 2m-temperature over NYC on the morning of 2020-01-01. The library # pushes WHERE clauses on dimension columns down to partition pruning. ctx.sql(''' SELECT AVG("2m_temperature") - 273.15 AS avg_c FROM era5.surface WHERE time BETWEEN TIMESTAMP '2020-01-01' AND TIMESTAMP '2020-01-01 05:00:00' AND latitude BETWEEN 39 AND 40 AND longitude BETWEEN 286 AND 287 -- ERA5 uses 0-360 longitudes ''').to_pandas() # avg_c # 0 8.640069 # Average temperature per pressure level, globally. ctx.sql(''' SELECT level, AVG(temperature) - 273.15 AS avg_c FROM era5.atmosphere WHERE time BETWEEN TIMESTAMP '2020-01-01' AND TIMESTAMP '2020-01-01 05:00:00' GROUP BY level ORDER BY level DESC ''').to_pandas()$

060

Al Merose @al.merose.com · 08/03/2026

Installing NixOS on my old tower computer. Might host a new stateful agent later

090

Al Merose @al.merose.com · 28/02/2026

First time I had a session with Claude that had a natural wrap up.

Everything is on autopilot for the night. Is there anything else you'd like to
work on, or shall we wrap up this session?

140

Al Merose @al.merose.com · 05/02/2026

I’m not through all the results yet, but I really like this paper. arxiv.org/html/2506.01...

4 Experiments
We now apply our method to analyze how embedding spaces evolve during LLM-related learning tasks. Unless stated otherwise, all results pertain to the model's final hidden layer, i.e., T_ı as defined in Section"3.
We refer the reader to Section"C.2 for results on layers other than the last. Subsequently, we will focus on four central questions:
(Q1) How does fine-tuning on different datasets al-
ter latent space geometry?
(Q2) How can local dimension estimates detect
grokking?
(Q3) How can local dimension estimates detect the
limit of training capabilities?
(Q4) How can local dimension estimates detect
overfitting?

010

Al Merose @al.merose.com · 04/02/2026

These are some of my favorite hungry ghosts

ANUPONG CHANTORN
Animal-Man Family
Bronze, 2010
Bangkok Contemporary Art Museum

100

Al Merose @al.merose.com · 01/01/2026

Nothing said “goodbye 2025” like the winery next door catching fire last night

020

Al Merose @al.merose.com · 01/01/2026

For NYE, I recommend “Al’s Sparkling Americano”: - 1 pt Campari - 2 pts Vermouth - combine and stir with ice - fill halfway with tonic water - top off with a double shot of espresso Tonight, I’m going with this nice Spanish vermouth, it’s nice and floral. Cheers 🥂

000

Al Merose @al.merose.com · 22/12/2025

A closing argument for sour cream over apple sauce in the great hannukah debate: dill.

The last latke from our batch, with sour cream and a heaping pile of dill. On kind of a messy plate.

100

Al Merose @al.merose.com · 20/12/2025

Christmas came early this year

A picture of the book by ARTHUR GILL:
“MACHINE AND ASSEMBLY LANGUAGE PROGRAMMING OF THE PDP-II”

020

Al Merose @al.merose.com · 15/12/2025

1: huh, they only have two of the three Abrahamic religions 2: oh, no, wait, the last one is just across the street

A sign that says “Los Alamos Perch, est 2017”. On top of the antique shop is a lit up cross and star of David (the six pointed star is above the cross).

The shop front of a bar called “Bar Alamo”. The bar sign features a crescent moon and stars. A horse is in the middle of the crescent moon.

010

Al Merose @al.merose.com · 27/11/2025

Happy Thanksgiving from Monterey

060

Al Merose @al.merose.com · 25/11/2025

I’d like to introduce an experiment I’ve been working on for the past couple of years: Xarray-SQL. This library asks, what if SQL worked natively with arrays? github.com/alxmrs/xarra...

Code demoing Xarray-SQL. It reads:
```
import xarray as xr
import xarray_sql as xql

ds = xr.tutorial.open_dataset('air_temperature')

# The same as a dask-sql Context; i.e. an Apache DataFusion Context.
ctx = xql.XarrayContext()
ctx.from_dataset('air', ds, chunks=dict(time=24)) # the dataset needs to be chunked!

result = ctx.sql('''
SELECT
"lat", "lon", AVG("air") as air_total
FROM
"air"
GROUP BY
"lat", "lon"
''')

```

172

Al Merose @al.merose.com · 15/11/2025

Call me picky, but I this this experience could be greatly improved.

100

Al Merose @al.merose.com · 05/11/2025

This is the best figure I’ve seen in a whitepaper in a long time, bravo @ai2.bsky.social, delightful.

Figure 2 Global distribution of data for OlmoEarth pretraining. We randomly sample 285,288 locations based on OpenStreetMap categories. What's your favorite map projection? I like the Peirce quincuncial projection centered on Antarctica. They said it didn't make any sense for this figure though. They said "We hate Antarctica, that's why we don't sample any points there". I said, "Hey, there is one point there!" They didn't have a response to that so they tried to silenced me by deactivating my acc

1104

Al Merose @al.merose.com · 19/10/2025

Sunday afternoons are for jazz and installing linux on your old laptops.

The Ubuntu 24 boot screen, which my reflection in the black of the screen

Spotify screenshot:

Alone Together
Kenny Dorham

150