LenaRc !🐀 (@Lenarcv1) header
LenaRc !🐀 (@Lenarcv1) profile picture

nooit meer slapen

Joined May 2022
Got really inspired after a short exchange with @LeahLundqvist a few days ago to work on my "infinite canvas" wayland compositor, and I also added a cute little agent that can work alongside you on the canvas through the IPC :O
6
1
27
3,909
LenaRc !🐀 retweeted
fuck woke 2 we’re going straight to woke 5.5 opus
63
1,588
43
29,534
465,893
Finally reading through moonshot’s PARL paper and to me it resembles traditional hierarchical approaches to this really closely, and Im so curious if putting the main optimization pressure on the orchestrator might cause different behavior in swarms compared to decentralized MARL
3
85
The world you grew up in no longer exists
1
13
169
.oO ( living as/in a jet..? )
Becoming a pilot nearly killed me, twice. It also saved my life. Fifteen years ago, my world was upside down. I was in founder mode building a company, raising three young kids, navigating a struggling relationship, leaving my religion, inflamed, sleep deprived, fatigued, and fighting chronic depression. Driving home from work one day, I saw a sign inviting prospective students to take a discovery flight. I thought this flight might occupy my mind, giving me an hour of desperately needed relief. Thirty minutes after arriving at the airfield, the instructor and I were seated in a Cessna 172. The instructor told me to take the controls and explained the departure procedure. Caught off guard, I suggested it would be better if I just observed. He insisted I take the controls, reassuring me that his hands would be on the dual controls at all times so not to worry. Reluctantly, I took the controls. We radioed our departure to the local airspace, lined up, and accelerated down the runway. At 63 mph, I pulled back the controls to raise the nose, and up we went. All of my life’s pain disappeared as I became absorbed in managing the throttle, pitch, heading, radio, traffic surveillance, and my instrument panel. I was mesmerized. I earned my pilot’s license two months later and bought my first airplane, a Bonanza A36. It could fly faster, further and have more passengers than my trainer airplane. I made myself a promise to never fly alone, and hired a full-time co-pilot. General aviation accident rates are nearly 10x higher than commercial airlines. I began flying myself everywhere across the country. For business: Austin, NYC, and Omaha. For family trips: Niagara Falls, Mount Rushmore, Bahamas, and later, with a longer range airplane, we flew to Europe via Canada, Greenland, Iceland, Scotland, London, Paris and Rome. In my first year of flying, I nearly died twice. I had hired a co-pilot for my safety, and ironically both incidents resulted from his errors. However, the worst mistake was my failure to recognize the bigger pattern: despite my co-pilot having logged over 1,000 flying hours, nearly all of the hours were “in the pattern”, meaning instruction time around the airport and not “in the wild” (out in high stress, cross country operations with severe weather). Lesson learned. After these two near-death experiences, I decided that I’d need to either stop flying or get a jet. A jet would give me engine redundancy and the capability to cruise above weather systems rather than through them. The Bonanza had a ceiling of 14,000 (before I needed to supplement oxygen), where a jet could fly in the 30,000’s and 40,000s, above weather. A jet would also allow us to always fly with the assistance of air traffic control in controlled airspace versus in un-cleared VFR (visual flight rules) flight. I purchased a small jet, a Citation Mustang, and completed my type rating (licensed to fly the airplane). It was a steep learning curve. Jet speed is different from small airplane speed. Everything happened faster, the airplane was more complicated, and multi-engine dynamics were a big step up. It was exhilarating. Part II In an unexpected twist, learning to fly brought me back from the depressive dead. When I started flight instruction, my personal life was a mess. The most irritating part of my existence was that every day, around 7pm, just after putting the kids down, I’d overeat. I tried everything to stop myself but nothing would work. My willpower at that time of day from the stress and strain of life was zero and I was powerless against the chemicals in my body that demanded this daily self-destruction ritual. My pants were too tight and I couldn’t button the top button. It drove me insane. But I refused to concede and buy a bigger pair. I turned to my pilot training to help me design a solution: I needed structured operations for my food consumption and self control. Aviation is substantially more safe than driving because it operates on structured protocols, checklists, and active risk management. Safety and performance are a system. You don’t just get into an airplane and wing it. Before every flight, there is a pre-check routine where you thoroughly inspect the airplane, ensuring everything is in flight-ready condition. There are other checklists for starting the airplane, taxing, taking off, cruise, descent and landing. There is a manual beneath your seat with a written protocol for anything that could go wrong with the airplane. You are required to memorize and execute procedures flawlessly under pressure. Aviation culture assumes human fallibility and designs systems to overcome those limitations and liabilities. As I was learning this new way of thinking and operating, I wondered if these new autonomous systems would map to my personal life. I was in desperate need of help and nothing I had tried was working. It was clear that Evening Bryan, the version of me who took control of my body at 7 pm, was no match for the stress and weight of the day. He wasn’t a bad guy, he just wasn’t equal to the responsibility of withstanding that tension. It was a system error to ask him to operate Evening Bryan under his duress. I put Morning Bryan in charge instead and he made an operational guide for all other Bryans to follow. Morning Bryan is my best Bryan. He’s responsible, motivated, high on will power, and thinks long term. He created a checklist that required all Bryan’s to be finished eating by 5 pm. No matter what, no more food after 5 pm. That was when the danger zone started. There wouldn’t be any exceptions, ever. This worked. It was challenging for the first few days but it quickly generated it’s own momentum. I started sleeping better, losing weight, and feeling less fatigued. That simple trick fixed years of mayhem that had been unfixable. With that success behind me, I expanded the protocols to include food types, amounts and timing. In retrospect, after a few months of this new system, it seemed weird that I previously winged it daily, stumbling from meal to meal. I felt like I got back 40% of my total energy capacity that was being spent thinking about, planning, debating, and regretting my food consumption. My life today basically runs like an airplane. Part III Aviation culture then jumped into other areas of my life. I’d started a new company, Kernel, to build a non-invasive brain interface to merge human cognition with AI. I’d been thinking about what mathematician Alfred North Whitehead observed: "Civilization advances by extending the number of important operations which we can perform without thinking about them." Society is on a continuous curve of increased automation. We solve one layer and step up on the next. This is how we scaffold progress. Clean water infrastructure, power grids, washing machines, dishwashers, and are examples of automating tasks that previously took hours. It was the mashup of aviation culture, Kernel, AI and automation that got me imagining building my Autonomous Self. An automated system that keeps me in pristine physical condition, youthful vibrancy and peak performance. A system that would allow me to not die. Don’t Die was born. For the past five years, I’ve been engineering the v1 human prototype for Don’t Die: an autonomous system built to preserve and protect life. It accounts for biological fallibility, offloading human limits to deterministic protocols. As we integrate deeply with AI, stepping up through layer by layer physiological automation will allow us to navigate frontiers previously unimagined. Immortalism is the corresponding philosophy that honors existence as our species' highest value and pursues that objective through quantitative and mathematical means. I never would have guessed that a discovery flight would have flown me in this direction.
1
1
4
561
Why is everyone acting like berghain is the epitome of a cool place to party that place is lame you guys are lame you dont want berghain you want to go clubbing and be able post it on social media. No better than the people at the Ibiza Anyma sets, just more status seeking.
2
135
I was banned from an EDH playgroup because I had this guy in my deck and he's evil in-lore. By extension I was labeled an eugenicist. Commander players are so cool
💧 urza💧 disponible en mi web #mtg #mtgproxies
7
123
Konami doing 42 hits just to put us in angelechy tier 0
1
2
110
LenaRc !🐀 retweeted
This is interesting work, but I don't think it's a significant update to the big picture of interpretability. I'm on board with the broad idea of doing exploratory interpretability with neurons, e.g. neuron-based attribution graphs, for pragmatic problems like why did a model do a sketchy thing? I think this work provides some reasonable evidence, in addition to good earlier work from Transluce, that this might actually work for practical purposes, and has some advantages over the standard alternatives like SAEs/transcoders. I do not agree with some people's grander inferences like that this is substantial progress towards ambitious reverse engineering, solving interpretability etc. I expect that a neuron-based attribution graph will still have significant error and be missing a lot just like transcoders, once you prune them down to the most important components, even if they solve error terms. Which is fine for pragmatic purposes, and may suffice for getting the gist of what's going on, but I remain pessimistic about really fully understanding things and being confident there's nothing we're missing. I do think that using neurons is just vastly easier to, for example, do on a frontier model, which is great. I'm also excited to see more innovation in using agents to interpret neurons. I feel like this is an area the field has somewhat neglected, and where there's probably a very high ceiling on how far things can go. The SwiGLU double negative thing is also a useful point. I still think superposition is real and this work doesn’t really affect my cruxes. Most importantly, it seems highly likely to be true that the model can represent more concepts in the residual stream than it has dimensions, ie representational superposition, which I think the authors agree with me on. For neuron/computational superposition, it seems to me like language models have far more facts and memorized things in them than they have neurons. And that superposition is an efficient way to pack in more facts than you have neurons by representing them as linear combinations of neurons, especially for arbitrary unstructured values. Facts/memorized content is largely an unstructured database of key-value pairs IMO and I don’t see how this is compatible with a monosemantic view of neurons. I expect that weak neuron activations contain lots of semantic superposition-y info, which, if ablated, would meaningfully damage network performance. As I understand the work, it does not properly address what's going on with weak neuron activations in LLMs (I am skeptical of how much vision results will transfer, eg I expect small image models memorise far less). This picture is compatible with the highest activations of a neuron being mono semantic (though IMO the post defines monosemantic too loosely, I think I’d call a fair few of those polysemantic). It’s also plausible that the most important concepts are likely to correspond to a neuron (or a handful), which would be great, and may suffice for practical purposes! Similarly, my impression is that SAEs generally don't find things that are a single neuron, or even very sparse in the neuron basis, and are often a better basis for certain tasks than the neuron basis, though the evidence base is kinda weak here since we largely focus on residual stream or layer output SAEs, I’d love to see work on this in Goodfire’s forthcoming paper! I also think residual stream SAEs are extremely useful, as they give a good bottleneck of the model’s information at this token and layer, and I don’t see a good way to replicate them in the neuron framework. And kudos to Nick for updating a bunch of his historical vision work! Revisiting old results with modern tools and more rigour seems great.
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works already-clean.goodfire.com/
39
175
29
1,762
136,034
Obviously superpositions are still a problem, they are super after all. Otherwise they’d be called normalpositions..,
9
225
I peaked at 7 years old. I peaked at 2600 meters in the swiss alps. I peaked at the top of celeste mountain in the hit game celeste. I peaked at gold in call of duty mobile. I peaked at high school because all the other schools are lower than the high one. I peaked when u touch-
1
9
136
I wish i was event food and people would like me even though im not really worth what they put in and they know it but they dont acknowledge the cognitive dissonance while they nibble on me send tweet
1
1
8
192
Can I have the xxxtentacion snare. Dude do you have the SOPHIE snare? Lmao this guy sampled the jack U snare where did he even get that sample. Yeah I used ai stem seperation to get the bnyx snare from a drake beat.
1
61
Wow Anne Carson so bad right. Yeah so Tumblr. Yeah. Good prose is when like, woof woof wruff. Nobel prize for writing but no barks have been found on page. Curious!
3
73
Got jumpscared today by someone I had never met before saying “Fuck my Labubu life”. Which they apperently got from a friend that got it from me.
18
I make fun of people playing league but grinding DuelingBook rated is probably at least as much of a pathetic experience so I’m a hypocrite
6
141
I thought everything in life was vagueposting but in reality I just needed glasses
10
143
I have like 5 oomfs doing “swarmchasing” atp and I haven’t looked into it yet, can anyone explain? Do you just scrape random forums & messageboards and try looking for weird activity. Or do swarms cluster around sites relevant to the task they do.
2
18
447
Uhm my caretaker just offered me a beer… I’m 87 :/
1
15
155
LenaRc !🐀 retweeted
I know you are all waiting for my opinion on whether the Genderbent Draco Malfoy is aspirational femininity designed to ignite revolution among the undersexed lumpenproletariat or necropornographic swill designed to self-smother the nascent male uprising. Thank you for waiting.
49
73
5
1,733
39,338