HNHacker News
TopNewBestAskShowJobs

minimaltom

240 karma · joined January 13, 2026

submissionscomments
minimaltom··on Qwen 3.8 27B
Architecture thread! Afaict they continue to use gated attention + delta net, which was also adopted+adapted by K3, but im surprised theres no improvements to the residual stream (deepseek are using manifold hyper-connections, kimi have attention residuals) ?

Perf improvements seem to all come from training?

minimaltom··on Gloomberb
I mean look at the name... this seems more like a fun quip than a product. I'm over here basking in the glow of dem vibes.
minimaltom··on Breaking the WAL
Yeah 100%! And I'm sorry if I sound a little more critical and less eager, its just thats theres a world of difference between a priori finding the bug, and reproducing it, and the impression of the article (from my read) was the former.

But please keep writing, I know its super hard to put yourself out there and make content!

minimaltom··on Breaking the WAL
Thanks for clarifying! It would be really interesting if Antithesis finds the bug when:

1. The specific bug isnt mentioned 2. (If youre game) a model with a knowledge-cutoff date before the report is used

minimaltom··on Breaking the WAL
I went clicking through to see if I could find the prompt they fed the AI to locate the issue / write the test suite.

I couldn't find it, so its unclear if the prompt was completely "make a test suite" or was lead towards finding it in the first place, which wouldn't be a fair test.

The closest I mention of the prompt I could find was:

> Then I asked it to write a simple workload which exercised the WAL insert and checkpoint code. Notably, this is a completely generic workload.

With a skeptical lens, unclear.

minimaltom··on Tailscale Traces Database Corruption to 16y/o SQLite WAL-Reset Bug
I'm equal parts intrigued and skeptical- I guess if the prompt doesn't lead on there is a bug there then I'm impressed.
minimaltom··on OpenSSH 10.5/10.5p1
Idk if its accurate to rote project the policies of OpenBSD to OpenSSH, yes technically its a subproject but in practice stewardship and thus effective policy is pretty much all damien.
minimaltom··on Show HN: Needle2: 14MB agentic LLM for phones, wearables, smart home and robots
What about mHC? I'm surprised it helped with such a small compute budget.
minimaltom··on Show HN: Needle2: 14MB agentic LLM for phones, wearables, smart home and robots
Was really cool to see yous use Engrams to cut down compute!

Given its basically an O(1) lookup with disk space being the main constraint, I was curious if you've tried ablating engram layers and sizes across your setup?

Also, why mHC over attention residuals?

minimaltom··on Study links GLP-1 drugs to bigger jump in women's employment than a degree
Sure the thermodynamic reduction of `weight-loss := energy-in < energy-burned` is true, but youre erroneously extending that reduction to assert basically ~variance in outcomes is explained by effort and measurement error~.

Most pointedly, theres a number of studies that measured outcomes from over/under-feeding identical twins [1] or varying exercise in identical twins [2] and found significantly wider variance WITHIN twin pairs than across the whole study. Phrased differently, calorie intake + exercise doesn't predict the outcome of weight even when controlling for hereditary factors, and looking at the geometric mean masks this.

TLDR the same amount of exercise + number of calories in one person doesnt predict the outcome in another. Hence why at-home delivery services doesnt eliminate the need for GLP-1 drugs.

[1]: n=12 twin pairs, Bouchard et al., NEJM 1990. [2]: "Mirror image experiment", Bouchard et al., Obesity Research 1994.

minimaltom··on Study links GLP-1 drugs to bigger jump in women's employment than a degree
I think this is a commonly-held misconception. Even when individuals accurately account for calories and workout intensity (using hard data, not feels), theres a large variance across individuals, broadly attributable to 'metabolics'.
minimaltom··on Study links GLP-1 drugs to bigger jump in women's employment than a degree
Usually, chronically-overweight ppl have tried regulating weight with all kinds of diets, lifecycle interventions, regular exercise, etc, and a lot of the time the thing that actually worked was the GLP-1. I think this speaks a lot to the variance in metabolism, hunger, cravings etc across the population.
minimaltom··on Zapscape (CVE-2026-64561): Guest-to-Host Escape in KVM/x86
Right, but what x86 KVM setup is there in practice that doesnt present a mmu to its guest and hence keep its own track (shadow) of memory mappings?
minimaltom··on Zapscape (CVE-2026-64561)
Yeah most cloud providers have live migration and a reasonably-well-tested playbook for updating the node kernel + hypervisor.
minimaltom··on Zapscape (CVE-2026-64561): Guest-to-Host Escape in KVM/x86
Oh yay another one lol. This one seems much more general than the prior one that needed nested page tables.

Patch Thursday for cloud VM ppl lol

minimaltom··on AI's debt binge can't last, hidden borrowing reaches $1.65T
Me too / born in Sydney! A good mental model is that its like super, the kinds of things that super can invest in are equally diverse and it all depends on who you setup with

Systemically i guess thats more a question of 1) is US equities just an AI trade, and 2) how much of all these investment vehicles are just such assets. Regardless of the answers, the nice thing is that unlike debt, there isnt a positive feedback loop here (i.e. margin calls increase volatility)

minimaltom··on Settlement with OpenAI for Discriminating Against U.S. Workers
This is actually nice to see, because sneaky things to maximize good outcomes for PERM like advertising roles on the radio late at night / in random community newspapers is (or was) standard practice.
minimaltom··on AI's debt binge can't last, hidden borrowing reaches $1.65T
"401(k)" is a class of tax-advantaged account for retirement, theres no one global 401(k). If you are concerned, you can look at your disclosures or ask your provider.

All of this is kind of beside the point though, because the issue was leverage not exposure. In south korea ppl were forced to sell at the low point of the market due to margin calls. For a retirement account, you can just choose to take a disbursement next month or next year (assuming youve managed personal cash flow with sequence of returns risk in mind).

minimaltom··on Boarding China's Last Bus
This was an incredible read, take my upvote!
minimaltom··on AI's debt binge can't last, hidden borrowing reaches $1.65T
South Korea had an extremely large population of retail investors investing in options and leveraged ETFs, to the point that 3% of the adult population has now been margin-called.

That setup isnt true for the US, not even close.

minimaltom··on ESP32-C3 SuperMini antenna modification
Paperclips can be very difficult to obtain sometimes :P (have you seen smiling friends?)
minimaltom··on ESP32-C3 SuperMini antenna modification
Hell yeah!!
minimaltom··on ESP32-C3 SuperMini antenna modification
Silver-coated wire? I don't have that lying around.
minimaltom··on Claude: Elevated errors across all models
Unless the load moving from down providers overloads the remaining ones, cascading into a catastrophic, model-wide failure.

(Luv a lil cascade, as a treat)

minimaltom··on After the AI Crash
Most new tech has insane P/E multiples for a time, for me the litmus test between "capex is high but but within a few multiples" and "capex is insanely overbuilt, ahh collapse etc" is whether revenue growth continues its trend line. I appreciate this is probably giving more grace to the bulls than its worth, but as they say, the market can remain irrational longer than you can stay solvent.

Btw the $2T/year number is just wrong, thats higher than current capex.

minimaltom··on Show HN: How far do I have to go to run into 100k people?
Even if you are colliding with 20 people a minute thats still 3.5 straight days of running into people. Even if gentle thats probably heavy bruising, and any conflict or discourse is gonna slow down that 20-people rate.

I imagine theres probably only a few cities where you could pull this off.

Anyways I know the article isnt this literal, this was just my first thought from the title and I had a chuckle lol.

minimaltom··on Kimi-K3 on HuggingFace
Thanks! I had a clanker graph it with sliders, its a very interesting function to play with! seems like it has the negative 'hump' typical of a GELU coupled with the saturation of a sigmoid.

Its interesting that saturation is ideal at this scale, i thought we were all in on self-activating functions (i.e. swish) but in fairness all those papers lacked ablations at scale nor discussion of stability.

minimaltom··on Kimi-K3 on HuggingFace
HF says activation function is "SiTU-GLU", but I can't find any info on that?

Anyone know / is this a typo?

Even the python code for inference seems to use normal activations.

minimaltom··on A solid-state “atomic channel” for separating rare earth elements
This seems very similar to the ion-exchange membrane rowow et all have been refining on youtube over the last year: https://www.youtube.com/watch?v=luulTI1RKHE&t=114s. Rowow himself has videos on using electrochemistry to extract rare elements from mining waste.
minimaltom··on AI Companies Are Trying to Hide a Staggering Amount of Debt
Even if theres a massive drawdown it will recover in the medium term (and in the short term is a great buying opportunity).

For the people who are close/early to retirement and can't do that, well, they need to manage sequence of returns risk.

Edit: I think some ppl might interpret this as me being bullish on the SP500. I'm not, I'm bullish on everything evens out and returns to the mean.

← PreviousPage 2 of 4Next →