Welp if you're that 0.1% that is super important ie secure critical infrastructure & information, you should probably do your best to downplay yourself, to force them to hire you on a consultation basis to maximize your bargaining power in what you can extract from them for your labor. kek Glad to see that they are cutting though.

@astralmatrix @a_musingcat i like to believe it's either The Rhythm of the Night or What is Love. ◕ω◕

what do you think is the next scaling paradigms? • Memory solutions (true memory) • Latent &or multi modal reasoning? • iterative training & progressive skill building, perceptual ability grokking (CUA, concepts on top of that, 3d reasoning, other pattern reasoning, goal-objective intuitions, navigation esp through time or around collision awareness, skills from other abilities & skills) I wonder how skill acquisition is affected by scaling sizes... 🤔🧐 • other things?

what about training for regularized skills esp iteratively & progressively & perceptions to beat games rather than over fit like some massively deep RL network? For instance what are humans without intuition, ability to understand space, pattern finding, sense of space, sense of time, sense of locality, ect ? perhaps focusing on a set of skills sequentially learned, is far closer to generalizing at some point. :x We have a lot baked in from nat selection & evolution. There's reasons to suspect that humans are composed of many smaller skills, perceptions, sensations that work with each other. For instance try moving one hand rotation away from you, while rotating your other hand in the opposite direction towards you. One clockwise, the other counter clockwise. The reason it's difficult is because it breaks the illusion of generality in humans, as we rely on smaller, common muscle twitches & heuristics pre-learned. (You can do it esp after some practice--- you're hands will attempt to sync in the same rotation either both clockwise or both counterclock-wise instead of remaining in counter rotation to each other ) Hence with all of our skills & abilities we are or seem quite general. I think it's quite plausible various Attention Heads and concepts slowly generalize after a massive amount of training. It would be tedious, but I don't think you'd have to beat every game. Then add in things like continuous learning, and reasoning over other modalities or using latent reps and things could be quite interesting soon. Or using image gen/physics gen tokens during the inference instead of purely text tokens. Then rewarding good reasoning chains. Hmm for my next move should i move to... or rotate that shape image of tic tac toe game or image aumentation hmm ah yes/ or no... well what if... But tbh would be nice if we had true memory... instead of frozen models, but I don't think you nec need that, esp if planning is good enough, though it should help '>.> Physics of Skill Learning sequentially learned Where does In-context Learning Happen in Large Language Models layers 15-25 Multimodal reasoning: Imagine while Reasoning in Space Token Assorted: Mixing Latent and Text Tokens for Improved Language Model Reasoning other interesting papers on LLM learnings: knowledge mechanisms representation engineering attention heads procedural knowledge *interesting detecting memorization Let’s Verify Step by Step

yeah, the endpoint models avail today are pretty terribad. From planning esp over longer horizons (until very recently), to understanding perceptually what is on the screen and what to do about it or navigation issues--- they suck Try using one with screenshots in playing Pokemon blue or using one to take orders on a custom gui of a McDonalds Point of Sales system and you will see precisely how bad they are. There are a LOT of issues. Curious to see the agentic endpoints soon and how they evolve by the end of the year. Hopefully we'll have agentic benchmarks like baskets of 2d games, 3d games, 3d tasks, simulators, GUIs, ect

@kimmonismus uh for now it's fine. Not a good idea to allow OS agentic LLMs from China if tensions get worse. May I remind people that LLMs can be poisoned to behave to specific secret inputs--- hence you could in theory have trip codes to install interesting capabilities & behaviors though

@RuneX_ai @RaphaelDabadie @sama real :3 at 13:30 - 13:45 "How many people feel smarter than gpt-4..." and "the progress from feb 2025-feb2027 will feel more than the progress from feb 2023-feb2025" at 50:20 -50:30

i mean i don't care who gets blamed, im a centrist kek. Could in theory blame the previous administration not doing anything about it, when they saw the mammal infections were spiking, or the feed thing bc they like to regulate or w/e. All i got is the data, no idea how ya stop this trend :x gl hopefully disappears on its own i guess

in the US the feed can contain raw infected poultry. Poultry litter is acceptable to feed to cattle, while beef cattle likely graze. >contaminated feed, contaminated drink (bird droppings ect), migratory birds overhead. Doesn't likely spread easily outside of birds, except through more direct means usually atm. (map 1: purple is bird only detections, by county) Widespread- but likely been so for a while tho ---not that big a deal, been bad for a decade🤔 (map 2: wild mammal infections- becoming widespread but spread is fairly slow, likely prey & contaminated drink --- This jump to mammals is fairly new and seems to be something happening over the past couple years.) (map 3: apx 900+ Herds of cattle infections to date -trending in the wrong direction--- not just a democrat conspiracy :x )

yes, but more in a insipid encroachment manner... Bodes not well, as it's spreading to mammals more easily, and seems to be becoming everywhere. So spread is going in the wrong direction. The risk is: will it continue these indirect infections (contaminated feed, drink, preying) and will all these interactions give it the ability to transmit between humans more easily. Basically if you see humans dropping dead, and we lose control of it (contact tracing fails to stop it); then go buy some Oseltamivir and hopes its enough, and get ready for a round 2. Hard to know if the apparent CFR and actual CFR will be or how many mild cases there may already be (thus going undetected but less severe than it appears) or if there are no/few mild cases and it's just actually that deadly.

@gdb rip Darth 🫡🥺 try not to let it sting too much, and cherish the memories you had. Sorry to hear of your loss. Cat's are indeed awesome. ./nods solemnly

Singularity Next Stops & speculations: ◕ω◕ Q1 2025 •Computer Use Agentic endpoints (Agents 1.0) •GPT 4o full modalities out finally : image gen, sound files • O3 full release Q2 2025 •Orion aka GPT 5/o3 (and next gen models shortly after by the other labs) Hmm is o3, orion confirmed?🧐 • Models continue to get cheaper to allow for more use of agents; base model $1.500 / 1M input tokens; Mini $0.125 / 1M input tokens. • More Post training RL tools like similar to Vision Finetunes but for full training for sequential skills ie building more skills on top of CUA endpoints, skills by layers, building regularized skills reliant on prev skills/perceptions trained to stack Q3 2025 • Sora upgrades; other vid gens are quite decent by now esp at control & edits; Real time human gens fairly common those still reqs some good compute •Finetune Store (assuming sequential stackable skills and training that can be stacked or souped from specific models --- likely not poss now if current Finetunes are Lora based shallow augmentations of outputs). Trading trained GPTs & also Selling the training sets used for some skill (images w rated responses.) • Continuous learning 1.0 • China tech exports crackdown further & beginning of the Open Source crackdown esp w China in mind (may not be fully rational?) Q4 2025 • multi modal reasoning chains 1.0 • memory informed inferences/ true memory • Agents 2.0 (now have better longer term task coherence, some spatial navigation intuition, and a jump in overall usefulness • Dev endpoints for free, but low RPM | low RPD endpoints for building & testing become more common • Models continue to get cheaper to allow for more use of agents; Base model $1 / 1M token; Mini $0.025 / 1M input tokens But also largest discounts are in images, making using agents more realistic, as sending a set of images is pricey $$$ esp if they have to reason about it... • More Agentic benchmarks, pref baskets of Simulators, Guis, Websites, 2d games, 3d games, Point of Sales systems 2026+ •"AGI declared" w Model after Orion --but still weird gaps in what they can do or learn out of distribution. Likely as a means for further funding or OAI ipo? lol. • Biochem models • material science models • Labs lean in harder on Robot & embodied ai & devices • OS from china gets full ban, OS generally gets restricted down • China tensions escalate

you are correct sir big labs slow to drip out various skills and/or still perfecting certain abilities ie Computer Use Agentic endpoints Also usually not well trained for situations outside of turn based conversations, ie poor 3d spatial understanding, poor navigation esp through time, poor understanding of cause & effect esp of novel patterns or abstractions (so like playing a game or using novel GUIs) Frozen models w until recently very poor long term planning & coherence long term nor ability to stay on task continuously reliably. No memory, so every inference is stateless. Basically labs are sitting on stuff, better agents is likely next thing to slowly drip out

@a_musingcat false dichotomy Go forward & ignore the sign to be a cat ◕ω◕ The paths are mere suggestions to guide your behavior

i wonder if it's more jitters about uncertainty or Stockholm syndrome. 🧐🤔 Bc if it's jitters about the future would prob be best to handle it w a clear sense of direction, solutions to address concerns, & optimism, while if it explicit value from their work (ie camaraderie & work friends, sense their work is important, or averse to change) probably needs a different take; ie fake jerbs regulated in for the interim, or focus on family, friends, hobbyist workshops, church, travel/living life ect

No. So far, generally falls under parody or attention grabbing imo it'd have to be pretty bad imo like something dangerous, encouraging or threatening harm (esp in a realistic way), or hard -isms. Not: oh genetics are such that races may differ in ability or in health. We should be careful.... (races may be inaccurate and offensive to some but well within bounds) Yes, would fire: ___ is superior. ___ is inferior. ____ should be removed from the population. (implying a threat to others down the line) <-- and thought you believe it or want others to believe that narrative ie to profit from/sell a book at the cost of others.) but not a flippant one of comment per se--likely see outrage from public anyways. I would prob let a few slide, to then watch for... or question you to gather more info on your state of mind and intent.

@OpenAIDevs very cool 😎

@outbreakupdates

@1goodtern how frequent is testing elsewhere of cattle for h5n1? could it be a don't test unless an animal death occurs, wont find the data issue? 🙈🙉🙊