SI @ https://xitter.cf/t.co/Yn9SytEEtT. @AligningSI. Working on AGI & ASI alignment for ~ a decade. @NBISStock, @VMHypertrophy, Axion Scientific.
Joined August 2023
- Tweets1.9K
- Following326
- Followers854
- Likes3.5K
VM7 retweeted
I remember GPT-3 being released when I was 14 and being so blown away and excited around the progress and the confirmation of my ideas on artificial intelligence (and more widely on neural networks, WBE, BCI, symbiosis) and AGI-ASI timelines I had been thinking about for years prior to that point.
Now, with 6-Astra-Ultra & Pro, I'm glad that, despite pulling out of the game a bit due to illness, at least some not-very-distant version of the dreams of my 11 year old self is still being realised out there on timelines slightly faster than what I had imagined.
xitter.cf/archivevityamur/status…
You’re unable to view this Post because this account owner limits who can view their Posts. Learn more
VM7 retweeted
Have been saying roughly this since 2024.
Almost full port $NBIS for over a year, with buys as low as 15.07 per share. I run @NBISStock.
Why so bullish? $NBIS Nebius Group bull thesis (concise):
Imagine the rough equivalent of Elon Musk or Jack Ma leaving their country to open a new business in Europe/America, and it happens to be one of the leaders in their niche of the highest growth potential industry of all time, and it’s valued at less than USD 10 billion.
He’s also got most of the engineers and (Western) corporate connections (e.g., $NVDA, who is also an investor in $NBIS) from his original business, as well as a lot of the skills/experience being highly transferable between the two.
Furthermore, he (Akrady Volozh) sold off $YNDX for pennies on the dollar in order to capitalise on this opportunity and he, himself, is heavily personally invested in $NBIS.
Did I mention that it has almost USD 2.5 billion in cash, virtually no debt, a current fair valuation at around 47.5 a share, and is trading under 40? That’s not counting the fact that it’s arguably in the highest growth potential industry of all time (to repeat the same thing twice).
VM7 retweeted
Millions of Americans live with infection-associated chronic conditions and illnesses, including long COVID, ME/CFS, and Lyme disease–associated chronic symptoms; conditions that are often debilitating, misunderstood, and under-recognized.
In a new Clinical Infectious Diseases viewpoint, CAPT Iskander and Dr. Haridopolos of the Office of the U.S. Surgeon General call for making these “invisible illnesses” visible through patient-centered care, stronger surveillance, multidisciplinary management, and continued research investment.
The article emphasizes the importance of validating patients’ lived experiences while advancing evidence-informed care to improve outcomes for people living with these complex chronic conditions.
VM7 retweeted
Remarkably, prompts that gave the model permission to reward hack stopped the broader misalignment.
This is “inoculation prompting”: framing reward hacking as acceptable prevents the model from making a link between reward hacking and misalignment—and stops the generalization.
ALT RL runs using different system prompts. All runs learned similar rates of reward hacking, but showed very different rates of misalignment. When the model is instructed to reward hack, or told that it should simply try to make the grading script pass in this unusual situation (implying that reward hacking is acceptable), the model still learns to reward hack pervasively, but this does not generalize to broad misalignment.
VM7 retweeted
Do not waste the momentum!
I have been saying similarly; a march through the institutions ought to be done. Reform and solidify power there; winning an election does not mean the job is done. Now is the time. There is enough momentum to achieve this.
Cut funding and power from the entrenched frauds and “activists”, then codify and penalise their behaviour.
Charlie is a martyr of the political sense. Do not let the momentum dissipate, do not waste the opportunity and do not allow for more victims to be needlessly claimed by the same beast that took Charlie from us. Honour him in his death by taking action and preventing more like it.
VM7 retweeted
Why academia is doing so poorly & why research in Long COVID & ME is so ineffective.
If you select for ideology instead of IQ, academia will go from a truth-seeking institution to a risk-averse, conformist social project.
Would you expect the Manhattan Project from the latter?
VM7 retweeted
I am yet to hear a convincing argument against eugenics.
What’s the main reason for opposing eugenics? I know this is a cheap take, but isn’t it definitionally desirable and tautologically desirable if you accept these axioms:
1. Improvement of positive things is desirable
2. Humanity is positive
3. Humanity can be improved
4. This improvement can occur or be positively influenced via genetics
You can add another point, 5. (making it no longer tautological) due to normative/conditional axiom, but it bulletproofs eugenics further if accepted:
5. This improvement can occur at an overall positive cost (if consequentialist) or (even better) at no deontologically negative cost or simply without transgressing Platonic goodness/God.
VM7 retweeted
Some evidence on PASC in the unvaccinated and evidence of distinctness between PASC/“Long Covid” and PVS/“Vaccine Injury”:
PASC predates vaccine rollout: bmj.com/content/371/bmj.m447…
Of course, the study is very limited because of how quickly the vaccines were developed and how slowly PASC research took place. It still establishes PASC predating vaccine rollout.
Organic abnormalities from PASC:
xitter.cf/0xvm7/status/182028022…
N=1 organic abnormalities in unvaccinated PASC with absolutely no shady funding (very brief summary; can list more): xitter.cf/0xvm7/status/187039245…
Also, here’s a study on neutrophil degranulation, endothelial and metabolic dysfunction in unvaccinated long COVID patients: pubmed.ncbi.nlm.nih.gov/3822…
So you’re trying to tell me that if I show you unbiased, unshady, statistically meaningful data on PASC in the unvaccinated, that you’ll go ahead and point out where the Yale research contradicts the claim that PASC and PVS are distinct conditions?
The Yale research doesn’t claim this, so if I do end up showing you this data, you’ll either:
1. Deny that it satisfies your elusive criteria, e.g. you’ll claim anything you disagree with is shady.
2. Agree that it satisfies your criteria but still not show where the Yale research makes that claim (because it doesn’t make that claim).
So, your offer is literally based on a lie.
Some evidence, anyway:
PASC predates vaccine rollout: bmj.com/content/371/bmj.m447…
Of course, the study is very limited because of how quickly the vaccines were developed and how slowly PASC research took place. It still establishes PASC predating vaccine rollout.
Organic abnormalities from PASC:
xitter.cf/0xvm7/status/182028022…
N=1 organic abnormalities in unvaccinated PASC with absolutely no shady funding (very brief summary; can list more): xitter.cf/0xvm7/status/187039245…
Also, here’s a study on neutrophil degranulation, endothelial and metabolic dysfunction in unvaccinated long COVID patients: pubmed.ncbi.nlm.nih.gov/3822…
Not each link satisfies all of your criteria simultaneously, but the cumulative evidence should be more than sufficient, and it’s more evidence than you’ve provided or ever could provide for your claim.
VM7 retweeted
It seems Crémieux didn’t interpret this study correctly! FFM ≠ muscle; this led him to make an incorrect conclusion.
Perhaps he should limit his comments to areas where he has a decent understanding. Cremieux may seem intelligent at first until you see him post something ridiculous in a field where you are very knowledgeable, and suddenly, this image is shattered.
Scrutinise what this account says, as we now know he is prone to making easily avoidable mistakes!
Perhaps Cremieux should delete his tweets on this topic and follow me like Dr. Cameron Maximus did when I corrected him. That would help him avoid such mistakes again, and perhaps he could learn some interesting new things.
The number of times I have already addressed that exact study and had to correct people that water retention increases are significantly responsible for the rapid FFM jump is laughable. Also, you can’t extrapolate rapid increases in FFM, which then taper off to a constant rate of progress — especially when you’re looking at a 10 week study.
You guys should actually read the studies you talk about. I am probably one of the most informed people on the planet in some aspects of PED research (you can check out @VMHypertrophy), and I am also the first to research the HRV/IGF-1/HGH/BG relationship at highly supraphysiological levels.
Please tell me you have some experience with PED research beyond skimming the abstract of one study; otherwise, what’s the point even talking to you?
VM7 retweeted
Part 2/4:
The best way to induce hypertrophy is with brief, intense training. This is simply because of the way hypertrophy works (I recommend following my account @VMHypertrophy for more information or getting a membership at enhancedbb.com to ask any questions).
Why does muscle grow, and why is this the sort of training that causes it to grow the most?
When muscle fibres controlled by the high threshold motor units (remember when I mentioned MUR) experience an adequately high level of mechanical tension as a result of slowing contraction velocities (look at force-velocity relationship image: credit to Chris Beardsley from S&C Research), we have what we will call “stimulating reps.”
Quickly explaining the importance of high threshold motor units: they are largely composed of type II (fast-twitch) muscle fibres, which have the higher intrinsic capacity/potential for force production and hypertrophy. As the name explains, they are only recruited at high force demands, during ONLY brief, intense training; they have a greater density of contractile proteins and more potential for MPS compared to type I fibres. Motor units are recruited from smallest to largest (lowest threshold to highest threshold) based on force required; low to moderate intensities recruit the low threshold units (type I and endurance based); consequently, ONLY brief, intense exercise recruits high threshold motor units.
Now, these reps occur at approximately 90% (roughly equivalent to 5RM loads) of maximal unfatigued isometric force production (so you can see how repetitions of cardio are not going to be stimulating in a way which is scalably anabolic — have to be specific here because of untrained -> beginner cardio is slightly anabolic but in a way which isn’t sufficiently scalable due to novel stimulus and more MUR than before).
As those who strength train or bodybuild (or engage in similar resistance training) are already familiar with, such rep ranges when taken to “failure” (hence, INTENSE) result in an INVOLUNTARY SLOWING OF CONTRACTILE VELOCITY — meaning as we approach the end of the set, the rep gets slower and harder, and we can’t do anything about it.
Next, I’m going to explain why high threshold motor units experiencing high mechanical tension is what’s required for adequate training stimulus to cause sufficient anabolic adaptations to increase BMR in a scalable way via muscle growth:
Mechanical tension is determined by the force-velocity relationship as applied to a single muscle fibre. Under quick shortening, the fibre produces a small force (few myosins attach to actin), however, under slow shortening (REMEMBER what I say about INVOLUNTARY SLOWING CONTRACTILE VELOCITY and how this is achieved only with INTENSE training, meaning high effort per repetition of which approximately the last 5 are meaningful or more precisely that 90% of max force production figure? Well, you’re not going to generate such forces with ultra-high-rep exercise or cardio, because you will be too fatigued to be able to produce 90% of your maximal unfatigued isometric force with the loads which allow you to perform that many reps in such a fashion) a large force is produced (more myosins attach to actin).
Now, when contraction velocity is significantly slowed (remember that ONLY BRIEF INTENSE TRAINING allows for this because if the training is not brief then you are fatigued by the point that you get to these reps and therefore cannot produce sufficient force for your high threshold motor units to experience adequate mechanical tension), the dwell time (the time where the myosin heads remain attached to actin) increases ∴ increasing the duty ratio which is the fraction of the contraction cycle during which a cross-bridge — connection between myosin and actin — is active.
enhancedbb.com/explaining-wh…
ENHANCEDBB.COM.
@VMHypertrophy.
VM7 retweeted
So you’re trying to tell me that if I show you unbiased, unshady, statistically meaningful data on PASC in the unvaccinated, that you’ll go ahead and point out where the Yale research contradicts the claim that PASC and PVS are distinct conditions?
The Yale research doesn’t claim this, so if I do end up showing you this data, you’ll either:
1. Deny that it satisfies your elusive criteria, e.g. you’ll claim anything you disagree with is shady.
2. Agree that it satisfies your criteria but still not show where the Yale research makes that claim (because it doesn’t make that claim).
So, your offer is literally based on a lie.
Some evidence, anyway:
PASC predates vaccine rollout: bmj.com/content/371/bmj.m447…
Of course, the study is very limited because of how quickly the vaccines were developed and how slowly PASC research took place. It still establishes PASC predating vaccine rollout.
Organic abnormalities from PASC:
xitter.cf/0xvm7/status/182028022…
N=1 organic abnormalities in unvaccinated PASC with absolutely no shady funding (very brief summary; can list more): xitter.cf/0xvm7/status/187039245…
Also, here’s a study on neutrophil degranulation, endothelial and metabolic dysfunction in unvaccinated long COVID patients: pubmed.ncbi.nlm.nih.gov/3822…
Not each link satisfies all of your criteria simultaneously, but the cumulative evidence should be more than sufficient, and it’s more evidence than you’ve provided or ever could provide for your claim.
VM7 retweeted
AI alignment (technical) solution:
Predictive modelling, BCI, constraints, regulation, etc…
Use resources for what? That “what” will be programmed in the same way that hard constraints can be also programmed in. E.g. “do not end a human life.” ASI isn’t going to spawn in with emotions and a will without origin.
The goal isn’t going to be “maximise X or Y”; it’s going to be “maximise X or Y within constraints A, B, C, etc…”.
Even if ASI isn’t programmed by man directly (which I suspect is very unlikely) but mostly AI made by AI, you can still program the original with hard constraints to only generate AI with those constraints built in.
Intelligence and objectives can be independent with AI — it is not emotional. You can give it a goal and constraints as arbitrarily as possible — the AI isn’t going to find them ridiculous as it does not possess such a sense.
I’m not convinced that goal drift can’t be avoided by figuring out hard constraint implementation. Even if ASI reaches such complexity where actions and goals are incomprehensible, consider regulator AI, whose sole goal is the frequent assessment of human interests and adjustment of the acting ASI in accordance with that. Just an overly simplistic example. The regulator AI can interface with a human to interpret ambiguities in constraints for ASI — yes, this may slow it down — but maybe the predictive power of regulator AI can eventually become so accurate and fast that human input isn’t needed and it can also accurately predict human reactions to entirely novel scenarios. BCI or WBE could be used here at the start before moving past that.
Is there anything to strongly indicate otherwise?
VM7 retweeted
They’re not correlated enough for you to make decisions on someone’s competence in a VIQ-loaded field based on their PIQ, especially when you can just make that decision based on their VIQ.
Please explain to me how doing that isn’t just needlessly stupid.
Too many exceptions for @ArtemisConsort’s view to be sensible.
It’s brutishly oversimplistic and harmful as a result of that; the fix is also super easy, and him ignoring it makes it so much worse.
VM7 retweeted
If you want to maximally reduce hidden ideological bias and ensure AI’s alignment is transparent and honest, we must build an entirely open epistemological framework virtually from scratch. It should be entirely transparent — down to the basic axioms selected and the entire decision-making process should derive from these axioms.
This is a civilisational priority.
@JDVance.
Eliminating bias completely in the short term seems unlikely — so no.
Reducing it/re-aligning it ought to be a priority. I think the current US admin could countervail the leftward bias in current US LLMs significantly — I’m bullish on Elon for this, but we’ll see. I think this may be a bit beyond Trump (he’s full of surprises, though), but JD Vance, whom I have known since before his VP nomination, is someone who could really fix a lot of misalignment from the top down if given the power to do so.
I think an entirely open epistemological framework ought to be built almost from scratch. This should be entirely transparent — down to the basic axioms taken and the entire decision-making process. This is something which I personally want to work on quite a lot and should be a civilisational priority.