gavinray

I used Gemini 4 Argon (High) from the Android app with the original image and following prompt:

  > "Edit this image to color code the benchmarks scores using a gradient scheme that is immediately obvious"

Here's the direct Gemini chat link:

https://share.gemini.google/g6LC7spTH0cO

  > are they essentially holding back sort of unweighted parameters built in to the model architecture which they can then populate with new context during 'sleep' and then using the techniques outlined in the 'dream' section to reinforce...

Sort of. As the model is interacted with, short term "memories" are accumulated in a sort of buffer/cache. During "sleep" these "memories" are given a ranking/score, and the useful bits are integrated into the weights.

They say the weight update can be done by computing a transform where the end state is the old weights with the addition of the useful semantic knowledge contained in memories. Rather than just additively chucking the entire memory concepts into weights. That's pretty neat.

The subtext for the memory architecture diagram sums the whole thing up pretty tidily I think:

  > Input streams from the active phase are stored in short-term memory before entering the sleep cycle, which bifurcates into Consolidation (transferring fragile short-term tokens into expanded long-term parametric memory using low-rank structures and MoE) and Dreaming (autonomous self-improvement via synthetic routing with random experts and RL refinement).

Benchmark scores recolored for readability:

https://i.imgur.com/IDFGuJU.jpeg

Spending up to 4.9% of their wealth to avoid a 5% tax is the fiscally logical thing to do

Oof. Well, in any case: https://sci-hub.su/https://doi.org/10.1086/601720

  > Bio is one of our sparsest regions 

Ah, that explains the topic distribution on the page. Can't fault it if the data just isn't there.

It does seem like one of the best uses of AI -- digging through existing literature to make novel connections/ideas.

Curious if you've heard of Don Swanson's "Public Knowledge Problem" paper from the 1980's, which was roughly this line of thought?

I'm an engineer by trade, but spend quite a bit of time in pharmacology and molecular biology research in my free time.

One of the proposal papers caught my eye:

https://lacuna.tiptreesystems.com/hypothesis/why-receptor-hu...

  > " ... a ligand that binds tightly and stays bound a long time should produce a strong intracellular signal, a weak binder a faint one."

When a ligand binds to the ligand binding pocket, the contact points it makes with the proteins inside (called "cofactors") and the strength of each contact point (measured in van der Waals force) modulate the distinct transcriptional effect.

This is called "conformation": the shape it takes in the ligand binding pocket.

So it is not necessarily residence time and overall binding affinity that modulate transcriptional effect.

A very good example of this is the androgen Trestolone/MENT, which is an order of magnitude more potent than Testosterone but displays similar relative binding affinity for the androgen receptor:

https://en.wikipedia.org/wiki/Trestolone

This potency is due to the specific conformational shape it takes within the AR binding pocket.

Family has all the traditional cardiac symptom clusters (hypertension, elevated lipids/triglycerides, diabetes) and family history of heart attacks.

My biomarkers were good, but combined with family history and a past history of drug abuse (primarily stimulants), it's sort of a wonder it didn't happen sooner.

The MI itself was unusual. When they placed the stent in my heart, the coronary angiogram was able to image my arteries. I had almost no measurable stenosis in my arteries: a minor 30% in one inferior artery.

I was at the gym when it happened, the prevailing theory is that I had the one artery with vulnerable plaque and it ruptured from arterial pressure.

So on the one hand, it sucks that it happened at all. But on the other hand, I'll take "you didn't die and the rest of your heart looks clear" with no complaints.

This is a micro-implementation of what research calls a "Neural Computer", in which the entire UI itself is generative and driven by user actions.

Neural Computers provide the illusion you are interacting with an underlying operating system/application, but are actually a recurring (user input) -> (predicted pixel output) loop.

In theory, it would be impossible to tell a perfect "Neural" computer apart from a "real" one. Which is kind of a wild thought.

https://arxiv.org/abs/2604.06425

  > who already have a great estimator for 10-year heart disease mortality risk in the coronary calcium score

Heart attack at age 29 survivor here.

CAC scores only show calcified plaque, not vulnerable plaque, which is the sort likely to rupture and cause an event.

You can go get a scan and receive a CAC score of 0, walk away thinking you're perfectly healthy, and then die tomorrow because the scan couldn't detect the massive amounts of non-calcified plaque lining your arteries.

The only non-invasive way to measure vulnerable/soft plaque is via CCTA

See, for example, "Atherosclerotic Plaque in Patients with Zero Calcium Score at Coronary Computed Tomography Angiography"

It measured nearly 1 in 10 patients with a 0 CAC score as having vulnerable plaque:

https://pmc.ncbi.nlm.nih.gov/articles/PMC5967134/

  > "The frequency of atherosclerotic plaque in the coronary arteries in 367 patients with zero CS was 9.3% (34 individuals)"

I didn't leave anything off, this is a direct paste from OpenAI's docs

https://developers.openai.com/api/docs?lang=curl

The code examples from their own website show the use of cURL and this CLI are roughly as verbose. Not sure I understand the need for maintaining a CLI.

    openai responses create \
      --model gpt-6-astra \
      --input "Write a short bedtime story about a unicorn." \
      --raw-output \
      --transform 'output.#(type=="message").content.0.text'

Versus

    curl https://api.openai.com/v1/responses \
      -H "Content-Type: application/json" \
      -H "Authorization: Bearer $OPENAI_API_KEY" \
      -d '{
        "model": "gpt-6-astra",
        "input": "Write a short bedtime story about a unicorn."
      }'

  > Having kicked opiates cold turkey in the past, I felt nostalgic for three simple days of traditional nausea, sweatiness, and bone-deep ache.

I am one of few people stupid enough to have, at one point, been addicted to and subsequently withdrawn from, almost everything but alcohol.

There is no way to explain the hell on earth that is benzo/GABA-ergic drug withdrawal, and I would not wish it on my worst enemy. It is an order of magnitude more intense than opiate withdrawals, and the timeline is weeks/months rather than days.

Like the author, I managed to quit via a year-long Ashton protocol taper.

The irony is that benzos were the one class of drugs I wasn't abusing for recreation. I was using mega-doses of Etizolam to cope with PTSD flashbacks, and only discovered I had become physically addicted after flushing it all and winding up in the ER.

I am a highschool dropout, so I am perhaps not the best judge of this, but I later worked on-campus at a major university and befriended many students, most of whom struggled to find employment after graduation.

I would counter that education, at least in the USA, does not adequately prepare one for life, the real world, and effectively finding employment.

What ACTUALLY mattered in my life turned out to be:

1. Who you know, and by proxy, what kind of a person you are

2. The value of "just showing up", every day

3. How important personal relationships/networking are, and just how SMALL the world is for how vast its size

Knowledge and skills are things you can sort of just pick up as you go along, but being given any kind of opportunity in the first place is not quite so easy.

Smaller than Kokoro yet sounds just as good, impressive.

Makes a great option for mobile on-device or browser app TTS.

Best sounding local TTS engine I've heard by far is Kokoro

It's small enough to be run on modern phones realtime, I use it for listening to PDF's/EPUB as Audiobooks

https://huggingface.co/hexgrad/Kokoro-82M

  > because the wilderness is behind 3 confirmation screens

Some of the earliest social/life lessons I learned were through interactions in that game

Today's youth will never re-live the experience of having someone tell you to "follow them to hidden area" and then one-hitting you with a rune weapon the second you crossed into wildy + taking all your shit.

I found the YouTube video the other day and enjoyed watching it.

The agent pathing seemed to clip a bit through the sides of mountains at times though -- or maybe that is just a bug of GM camera following player?

Because telling lies on the internet attached to my real name is so beneficial to my reputation, right?

https://i.imgur.com/CNAUczi.png

I lost my ability to fully explain the systems I work on around the time I transitioned from next-edit-prediction AI running in my IDE (early Copilot) and working primarily via Codex/Claude CLI's.

The further away I got from the implementation details, the fuzzier the semantics became. To the point that now I could not in good faith say I understand the system I have built from the ground-up and worked on the last ~4-5 years.

The solution to this, in my mind, is not to revert back to reviewing implementation details. The economics of that from a business/productivity perspective don't play out well.

The only viable solution in my mind is adapting code/languages themselves to the shift by making programs formally verifiable. That way you can be sure that the code "does what it says on the tin", and provably so.

I feel so strongly about this I've written on it:

https://gavinray97.github.io/blog/design-by-contract-and-eff...

  > Self-improving code would be code that runs itself and improves itself.

I don't find the description any less accurate than saying that a JIT'ed application is self-improving code.

  > "... for the harnesses we have seen in these campaigns, average success rate is hovering around 10%"

A 10% success rate is absurd.

Just got an email that I now have access to Mythos:

  > "Your Cyber Verification Program access now includes Claude Mythos 5.1, Claude Opus 5.5, and Claude Sonnet 5.5."

This was a lot more exciting than I expected.

The homepage does not do a very good job of explaining what we're looking at:

It appears to be a combination of two things:

1) A formalization of the Python language

2) An LLM which hooks into running programs and hot-swaps code with verifiably-equivalent behavior but better runtime performance, found via superoptimization and then verified via the formalization rules

This is brilliant, and in hindsight one of those "I can't believe nobody had thought of this yet" sort of things.

Here's the transcription of the reasoning:

  > Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd. (hereinafter DeepSeek), a foreign, privately-owned company, located in the People’s Republic of China (PRC) provides multiple artificial intelligence applications available widely and free of charge. DeepSeek’s models present multiple threat vectors. DeepSeek servers are located in the PRC, and the PRC legal system permits extensive authority to government cyber and intelligence functions to access private company infrastructure. For DeepSeek-hosted instances of their models, DeepSeek’s account generation process collects personally identifiable information (PII) and may also track and digitally fingerprint users. DeepSeek’s most widely used applications function through a prompt-and-response process that requires presenting information to the model, and which may result in the transfer of sensitive data to the PRC if used by U.S. government employees or vendors. The transfer of the aforementioned information to the PRC, either independently or in combination with publicly available data, could produce actionable intelligence on U.S. persons and/or government programs. The models themselves may contain malware or other security vulnerabilities. Indeed, open-source security research has found that the DeepSeek iOS mobile application contains multiple critical vulnerabilities.

Kiwi and Boba (tea)

If you take this question far enough, you eventually reach a core philosophical debate:

"Does it matter what anything innately is, beyond the effects it has on you?"

For example: Suppose you were given a human companion with a robot brain that was your ideal friend. They made you laugh and you enjoyed every moment you spent with them.

Does it make it any less impactful/genuine to you that they're not a "real" human, or is your felt experience the only meaningful end-measure?

For me, it's the latter.

  > Spamming the government, using a sick mother as leverage and threatening the small claims court.

Here I am uncomfortable with telling the waiter I didn't want lettuce on my burger...

As another person commented, we currently believe the primary mechanism behind the pleitropic effects of GLP-1's to be a reduction in systemic inflammation and direct improvements in vascular/endothelial function.

The best paper I've seen on this was the recently published and very aptly named:

"Pleiotropic effects of GLP-1 receptor agonists — a narrative review of extra-metabolic outcomes in adults" (2026)

https://journals.viamedica.pl/medical_research_journal/artic...

Possibly the best and most succinct explanation in the thread thus far.