joshwa

calibrating the LPPLS model using Deep Learning: papers.ssrn.com/sol3/papers.cf…

i watched bugonia recently. clever film. i wrote out some thoughts below. a review, i suppose or the thoughts it evoked from me. the term bugonia refers to an ancient belief that bees could spontaneously generate from within the carcass of a cow. to a modern audience, this idea appears absurd. we are confident that bees do not originate from within a dead animal. yet the observable result remains the same: bees appear in the carcass. whether they moved in or spontaneously generated becomes, in some sense, an extraneous detail. reminds me of work by Simon DeDeo (@LaboratoryMinds) on how humans build causal narratives to explain complex systems. this idea becomes a useful metaphor for understanding the character of Teddy in bugonia. over the course of the film, Teddy lays out a set of theories that the audience is encouraged to see as irrational. the film reinforces this framing through moments like the dinner-table conversation about internet echo chambers and targeted online content, where anyone can find information that confirms their preexisting views. when Michelle tells Teddy that he needs help, the audience is likely meant to agree with her. Teddy appears less like a whistleblower than someone lost in conspiratorial thinking. but the metaphor of bugonia suggests a different way of interpreting Teddy's worldview. in bugonia, the explanation is wrong (bees do not spontaneously generate from a carcass) but the observation itself is not entirely mistaken. there are still bees in the carcass. Teddy's belief that Michelle is an alien may function in a similar way. by the end of the film, the alien question becomes almost secondary. what matters is that the CEO’s actions on earth have marginalized Teddy, his family, and potentially many others. the film repeatedly emphasizes the trauma that shapes Teddy's life: his mother’s participation in Auxolith’s failed and harmful drug trial undertaken for financial survival, and the abuse he suffered at the hands of his babysitter (who is now a police officer). in both cases, figures who held authority or power over him betrayed his trust in profound ways. Teddy's seemingly absurd belief system can therefore be understood as a response to these experiences. if the film’s recurring idea is that “in the end there are still bees,” then Teddy's parallel would be: in the end, there is still trauma. whether that trauma originates from an extraterrestrial infiltration of earth or from the actions of a powerful and unaccountable corporation becomes a comparatively minor distinction. the harm itself is real, regardless of the explanation Teddy constructs to understand it. seen in this light, bugonia is not merely a story about irrational belief. rather, it offers a critique of the social structures that produce marginalization in the first place. Teddy's worldview may appear delusional, but it emerges from genuine experiences of betrayal, exploitation, and institutional failure. in this way, bugonia suggests that what society dismisses as irrational belief may sometimes be a distorted response to very real harm.

@AIatMeta @NeurIPSConf @karpathy curious what you think.

@karpathy still bullish on token level standardization

@patrick_oshag love the melly andrew

I spent last night with Andrew Strominger and Alex Lupsasca, two of the top physicists in the world They just released a paper, co-authored with OpenAi, that seems to me like ASI Andrew, who helped develop string theory, told me that a year ago, his view was that he didn’t know how helpful AI was going to be. A year later, after some back and forth with GPT 5.2 pro, they submitted a final query to an internal model which solved AND proved a previously unsolved problem in quantum field theory…in 12 hours. A model, doing something two of the smartest people in the world in their field couldn’t do. And, when I was with them, they were giddy with excitement for what might lay ahead. Andy said “It is the first time I’ve seen AI solve a problem in my kind of theoretical physics that might not have been solvable by humans.” They said, “two things changed: the model improved and we figured out how to talk to it.” Andy also told me “I also now feel that with the recent advances, most physicists who want to keep up with the frontiers of progress will need to learn how to talk to it. That wasn’t true a year ago.” ASI is here, just not evenly distributed.

the Toolformer paper from @AIatMeta (ai.meta.com/research/publi…) present at @NeurIPSConf in 2023 showed how LLMs could use special tokens (reserved markers distinct from regular words, learn more: huggingface.co/learn/agents-c…) to invoke external tools/APIs. The pace since then has been rapid! Across major proprietary foundation models (@OpenAI, @GeminiApp @AnthropicAI) there is convergence on API implementation for their LLMs but an observed divergence in implementation of special tokens in their training data. This creates a dilemma for developers; should they relinquish context management to the LLM provider? Or, maintain control and invoke tools based on their best guess of the special tokens? Below are the two options reiterated: 1. Provider-Managed (API Abstraction): where the provider handle tool calling via APIs, abstracting the mechanism that intercepts tokens. 2. User/Model-Managed (Direct Token Control): Open models like @AIatMeta's Llama ([INST], see #L44" target="_blank" rel="nofollow noopener">github.com/meta-llama/lla…) & @MistralAI's models ([TOOL_CALLS], see #control-tokens" target="_blank" rel="nofollow noopener">docs.mistral.ai/guides/tokeniz…) expose signaling via explicit tokens. Wouldn't standardization at the token level (a shared markup language for training data) be beneficial, perhaps in addition to a protocol at the API layer? Many might prefer simpler completion endpoints & transparent control via standardized tokens. We see OSS projects like @openbb_finance's OpenBB Agents (github.com/OpenBB-finance…) & @cline's Cline (github.com/cline/cline) already opting to manage tool calling themselves. @AnthropicAI's Model Context Protocol (MCP) (modelcontextprotocol.io/introduction) is a positive step for standardizing connections (their "USB-C for context" analogy 👍), but it doesn't standardize the interaction language itself i.e., the specific tokens signaling LLM intent. I want this missing piece: a standard, reserved set of signaling tokens ([API_CALL], [ROUTE:TOOL_X], etc.) understood across models. This would enable direct control via completion endpoints. Then all the foundation model providers could just fine-tune next-gen models on these primitives. Hurdles def exist, primarily conflicts with provider business models built around function-calling API features. Documented access to existing special tokens could be a start? 🙏 This leads to a fundamental question as LLM workflows grow complex: Where should context & interaction management live? Will provider APIs remain primary, or will developers need the control of standardized, token-level protocols? What do you think? ❓ Prefer API abstractions or direct token/markup control? Why? ❓ Is the lack of standardized signaling tokens a bottleneck for you? ❓ Is token-level standardization even feasible? ❓ Where should context/interaction management ideally reside – our side or their side of the API? Let me know below! 👇 #LLM #AI #Toolformer #API #FunctionCalling #Standardization #DeveloperExperience #AIdev #FutureofAI #MCP #OpenSource

@__paleologo La storia è un cimitero di aristocrazie.

Just this week, I have listened to a podcast where the author (CS prof at U Toronto) predicted the end of democracy. The Google X business manager predicting the end of capitalism. Uncountable people predicting the end of white collar jobs, blue collar jobs, extermination of the rich. A number of them has respectable jobs. The population of forecasters is growing at an alarming rate. It is a very understandable reaction to uncertainty. But it doesn't make a lot of sense to predict the future when one a) has not practiced long-term forecasting; b) has not tracked/calibrated their performance; c) the future is coming faster at you thank you can predict it. On the other side, this helps explain because no one seems to agree on the valuation of anything. But can you all please go back to posting technical papers and pictures of cats.

@suchenzang curious what you think about this.

"The conscious and intelligent manipulation of the organized habits and opinions of the masses is an important element in democratic society. Those who manipulate this unseen mechanism of society constitute an invisible government which is the true ruling power of our country" -- Edward Bernays, "Propoganda" 1928

I want a YouTube channel that psychologically decomposes tv commercials.

@VictorTaelin Dali-esque

Gemini 3

a cat riding a horse opus 4.5 × gemini 3

@RobertoDailey1 the insight is promising for MAS. but it does trivialize task decomposition a bit. like, a step in towers of hanoi is always the same thing: move a disk. operationalizing task decomposition in a complex setting is not so trivial. but very interesting nonetheless. great work!

New work from Cognizant AI lab: Solving a Million-step LLM Task with Zero Errors. Existing LLMs struggle on long task horizons as persistent error rates compound, even when the LLMs know how to solve the task. Apple’s “Illusion of thinking” demonstrated that state of the art reasoning models could struggle with a simple task, Towers of Hanoi, if that task required execution of hundreds of steps in a row without error. We hypothesized we could see much higher performance by taking breaking down the task into its smallest subtasks, then using voting and red flagging to boost subtask accuracy. With these simple modifications we were able to push the simple gpt-4.1-mini to solve the 20-disk towers of Hanoi, or 1,048,575 steps without a single error! Seeing these results we believe with the right, robust, frameworks, LLMs can be scaled to vastly longer task lengths than their base model. Paper: arxiv.org/abs/2511.09030 Blog: cognizant.com/us/en/ai-lab/b…

@choffstein like kinda cringe?

sometimes I wonder if – at the time – led zeppelin's lotr reference would've been like imagine dragons referencing game of thrones.

do humans learn through gradient descent? prob not. does sequence modeling get you to agi? prob not. is sequence modeling pretty amazing? yes. the fact that you can get human intelligible text from next token prediction is astounding. i wish we had more commentary on this. how do wet networks work? where does the analogy with artificial neural networks breakdown? wet networks are undirected. ANNs are not, at least not in the same sense. forward pass, backward pass is not same as undirected. brains do use electric/chemical signaling across the neural network using synapse dendrites etc. curious to learn more about this.

@ColdCapital i sense sarcasm ex ante calling for 60 but then ex post it is at 65... was i misreading or is it a pie in face?

@BldrInvstTech What do you need help with?

i’m not sure i know what’s going on here. can someone help plz?

@richardcraib is this a commentary on happiness? maybe on purpose?

one of the craziest truths is you can start a restaurant or an ice cream shop or whatever and then work so hard it’s beyond belief how hard you worked and how much you sacrificed and then you double check your earnings 30 years later and you’ve made 1/3rd the s&p return… what was the point of your sacrifice? your efforts? you missing the school play? you wanted to help your community you say you want cheap ice cream for everyone and that was your sacrifice and that was how you gave back. okay so now what?

this is awesome @JayAlammar! the ELI5 is clutch. looking forward to @NeurIPSConf jalammar.github.io/assets/neurips… x.com/BldrInvstTech/…

sheet for data: docs.google.com/spreadsheets/d…

update after another ~3 months. still seeing steady growth. some commentary: i have been bearish on mcp because it doesn't have a solution for tool pollution. i think the pipe-dream for everyone is to have this chatbot that can access all their data from any system of record and synthesize it for good use. the fundamental limitation is the expressiveness of the tools available and the LLMs capability to compose those tools. an analogy from my prof that i really liked is this: "when an engineer sits down to write code that solves a problem, they don't write out the individual functions then compose those together. instead they do the opposite, they write the complete script to perform the solution then after the fact, they may refactor to composable units." so all that to say, this code execution direction from @AnthropicAI is compelling: anthropic.com/engineering/co… what's more expressive than a programming language?