Ross Taylor and Chengxi Taylor of General Reasoning discuss scaling reinforcement learning to long horizons, framing it as a mindset of staying coherent over extended agent interactions. Topics include Taylor's earlier work on Galactica with curated data and intermediate reasoning tokens, the role of value models in reducing variance and aiding credit assignment, bootstrapping signal from sparse rewards, the tradeoff between off-policy staleness and GPU utilization as rollouts lengthen, and a concrete task where frontier models traded football matches with real money and performed poorly, illustrating that long-horizon scaling depends on better environments and simulation rather than larger context windows. # Channel Identity Observations — General Reasoning Talk ## Explicit Self-Description (Creator Claims) - **Company identity**: Ross states, "I'm the CEO of GR. We're a London-based reinforcement learning company." Chengxi later repeats: "General Reasoning, we're London-based AI research company." - **Speaker credentials claimed**: Ross: "before GR, I was the reasoning lead at Meta AI working on Llamas, uh Galactica." Chengxi: "co-founder and president of General Intelligence Inc." *(note: "General Intelligence Inc." vs "General Reasoning" inconsistency in transcript).* - **Products/platforms claimed**: "Open Review was a product is a platform by General Reasoning... openreview.ai. So, it's a place where host over 350 environments and with a single API endpoint." - **Mission claim**: Chengxi: "long horizon is not just engineering problem. It is a mindset... if we really are ambitious to solve humanity's biggest problems, this is the journey for everyone. And that's also the mission for general reasoning." ## Claimed Scope of Content - Technical talk structure announced: "we're going to talk about algorithms, environments, compute, all the things you need to do to get agents scaling... to kind of long [horizon] tasks." - Topics demonstrated: RLHF history, Galactica, Llama post-training, RL with verifiable rewards, PPO/GRPO/critics, compaction, pipeline RL, GPU utilization trade-offs, the "Kelly bench" benchmark (Premier League trading), featured "in the front page of the Financial Times" (creator claim). - Promotional call-to-action: "if you are interested, follow us." ## Intended Audience - **Technical/research ML audience**: assumes familiarity with RLHF, PPO, GRPO, value models, context windows, off-policy trade-offs — presented without basic definitions. - **Conference/event setting**: "in today's conference room, there are other real people"; references to "many of you used Papers With Code back in the day." - **Potential customers/collaborators**: mentions the platform is used "for our internal RL and also some frontier labs and new labs are using this." ## Observable Presentation Patterns - **Dual-speaker founder talk format**: Ross covers historical/personal narrative; Chengxi covers forward-looking technical roadmap — an explicit handover structure. - **Blend of research narrative and product marketing**: technical content interleaved with platform promotion (openreview.ai) and benchmark publicity (Kelly bench, FT coverage). - **First-person insider framing**: heavy use of "we" for both Meta-era work and current company, e.g., "we were kind of tasked as a team to kind of spin up the post-training effort for Llama." - **Unpublished-claims style**: "we still had a recipe at the time, and this is unpublished, but it was really good for the meta" — claims offered without external verification in the transcript. ## Distinguishing Claims from Demonstration - Demonstrated: technical depth in RL reasoning, long-horizon RL discussion, environment-platform existence claim with named URL. - Claimed only (not independently verifiable in transcript): state-of-the-art results, FT front-page coverage, 350+ hosted environments, "billions" potential of the benchmark, insider credit for "the open weight kind of revolution." # Delivery Observation: Scaling to Long Horizons ## Overall Format - **Two-speaker structure**: Ross delivers a personal/technical narrative first, then hands over ("I'm going to hand over to Chengxi") for a forward-looking section. Closing summary explicitly merges both: "to summarize both Ross and my speech." - Talk framed as insider storytelling plus roadmap, ending with a company pitch ("if you're interested, follow us"). ## Ross's Delivery Patterns ### Heavy Verbal Fillers - Constant **"uh"** and **"um"** throughout: "Uh so this talk is called…", "Um so this was the Papers With Code team." - Frequent **"you know"** as connective tissue: "you know, the early days", "You know, they were both…" - Also **"like"** as a hedge: "It's where you prompt for like kind of like the steps." ### Recurring Phrase: "kind of" - Extremely dense repetition of "kind of" as a softener: "a different personal perspective", "this like crucial reinforcement learning…", "this kind of recipe", "this is like kind of what Deep Seek 01 ended up doing." ### Narrative Framing & Signposting - Explicitly announces structure: "we're going to have two parts of this talk today." - Story-announcement habit: "So let me tell you a story", "let's talk about Galactica." - **Lesson extraction habit** — repeatedly states morals of stories: > "that's like a good like kind of lesson there. A good base model is not enough." > "that again reinforces base good base models not enough." > "it was just like the bitter lesson, like the most purest form of bitter lesson possible." - Numbers delivered as punchy contrasts: "two orders of magnitude fewer parameters, but getting better results", "double the performance, order of magnitude less results." ### Self-Deprecating Informal Humor - "I was a very sad panda, uh defeated by ChatGPT and O1." - "we were focused on a, you know, bloody base model, which is hard work in 2022"; "Galactica was actually a bloody good model." - "It wasn't too fun." ### Addressing the Audience - Direct listener engagement: "I'm sure many of you used Papers With Code", "as you guys know now", "some of you might remember this." ## Chengxi's Delivery Patterns ### Distinct Filler and Grammar Style - "So" as an opener for nearly every point: "So one solution is use compaction", "So to solve this issue, we can apply critics." - Non-native article/verb patterns observable: "what it take for the mathematician", "there's a real money to be made", "it will goes beyond the constraint", "if it was a successful, couldn't make a billions." ### Repetitive Rhetorical Scaffolding - **"But here's the problem" / "But, here's the issue" / "here's another trade-off"** used repeatedly to pivot to challenges. - Enumerated listing habit: "the first is… And the second is… And finally…" - Analogy to biology: value bootstrapping "It's like a dopamine in human brain." ### Motivational Bookending - Opens and closes with the same thesis, verbatim-near repetition: > Opening: "Long horizon task is not just an engineering problem. It is a mindset." > Closing: "long horizon is not just engineering problem. It is a mindset." - Grand-scale framing: "cure cancer, solve millennium's prize problem, or go into Mars." ### Trade-off Framing - Consistently presents solutions with costs: "So, essentially, we made a trade-off between the off-policy and the GPU utilization… But, here's another trade-off." ## Shared Patterns - Both use **"So"** heavily as a discourse reset. - Both employ concrete benchmark numbers (Ross: "68% compared to GPT-3.5, 49%"; Chengxi: "100K", "350 environments", "eight steps"). - Both end sections with a forward-looking pitch for their company. ## Stability vs. Isolated Moments - **Stable**: Ross's fillers/"kind of", story-then-lesson structure; Chengxi's "here's the problem" pivots and mindset bookend (repeated within the talk). - **Isolated / lower confidence**: Ross's humor ("sad panda") appears once — cannot confirm as a stable habit from this single talk. Whether Chengxi's grand-mission framing is habitual is uncertain on one sample. # Position Record: Scaling to Long Horizons — Ross Taylor & Chengxi Taylor (General Reasoning) ## Speakers - **Ross Taylor** — CEO of GR (General Reasoning), London-based RL company; previously reasoning lead at Meta AI (Llama, Galactica). - **Chengxi Taylor** — Co-founder and President of General Reasoning. --- ## Part 1 — Ross Taylor: Historical Perspective ### Background - Started with **Papers With Code**, a London startup (2019) acquired by Meta; team then did Galactica and post-training for Llama 2 / Llama 3. - Claims the "open weight revolution" started with this small team; argues small aligned teams can achieve great results even in the scaling age. ### Galactica vs. ChatGPT — RL as the differentiator - Both were based on good base models; the key difference was ChatGPT's **RLHF pipeline**, which "made LLMs products for the first time." - Frames the two-week gap between Galactica's release and ChatGPT as a **"natural experiment"** showing RL provides value. - Stated lesson: **"A good base model is not enough."** ### Evidence that basic RL matters (pre-ChatGPT) - Cites **InstructGPT (2022)**: a 1B-parameter RLHF model outperforming 175B models — two orders of magnitude fewer parameters. - Admits his team wasn't paying close attention: "maybe we should have been as well, but we were focused on a bloody base model." ### Galactica's contributions - Claims Galactica outperformed PaLM, Chinchilla, and GPT-3.5 with less compute in scientific domains (e.g., ~68% vs GPT-3.5's 49% on LaTeX equations/science; 36 vs 19% chain-of-thought vs PaLM 540B with 30B parameters). - Key ideas he attributes to Galactica: - **Data efficiency**: 105B token curated corpus vs trillion-token corpora — contrarian at the time. - **Multi-epoch training**: first major empirical result challenging the "one epoch only" consensus (the ~4-epoch rule formalized later). - **Thinking tokens**: internal working memory within tags, spending inference compute before answering — a buried but important idea. - Acknowledges Galactica "blew up" publicly; the Meta association "didn't help"; novel work was overshadowed. ### Unpublished reasoning effort ("RSP") at Meta - His personal idea: apply RL pressure to the thinking-tag content. - Notes this resembles what **DeepSeek (R1)/O1 did ~2 years later**, but with key differences: - Only Llama 2 base models, poor math corpus, 4K context window. - Recipe: (1) continued pre-training on math/science; (2) **PPO with verifiable rewards** (not GRPO), with an outcome reward model initializing the value model. - Achieved internal SOTA math/reasoning results — but **no inference-time scaling or reflective behavior** (backtracking) that characterized R1/O1. ### Why no "moment" then? - Answer reached ~2 years later: **better base models** were what "got RL cooking" — framed as the purest form of the **bitter lesson**: better base models + more RL compute + bigger context windows yield emergent behavior. - Sociological observation: OpenAI having a GPT-4-level model first let them "see further" — scaling creates a research advantage. - Describes himself in 2024 as "a very sad panda, defeated by ChatGPT and O1," but still convinced reasoning wasn't solved — motivation for founding GR. --- ## Part 2 — Chengxi Taylor: Scaling to Long Horizons ### Core thesis - **"Long horizon task is not just an engineering problem. It is a mindset."** Solving humanity's biggest problems (cure cancer, Millennium Prize problems, Mars) requires patience and long-horizon thinking. ### The context window problem - Fermat's Last Theorem analogy: a mathematician's ~10 years of reading, scratchpads, and reflection ≈ tens to hundreds of billions of tokens — vs. current ~1M token contexts. - **Compaction** as a solution: generate until context limit, summarize, continue. RL can optimize both compaction and the task ("kill two birds with one stone"). ### Three RL problems with long horizons 1. Gradient variance scales with length. 2. Sparse rewards / credit assignment. 3. Variable trajectory lengths complicate optimization. - **Critics (value models)** reduce variance, fit compaction at the trajectory level, encourage batch diversity, and enable bootstrapping — but are more complicated than GRPO (must train a second model). ### Tools for context limitations - **File system tools** (scratchpad), **self-search tools** (search past trajectory), **archive tools** (build on previous results). - Caution: archives can let AI "cheat by just grabbing the previous answer without thinking." ### Kelly Bench benchmark - Agents build ML models to trade on Premier League football over a 1-year horizon; featured on the Financial Times front page (mainstream interest "caught us off guard"). - Motivation: real money, real-world implications unlike "cargo" (Kaggle-style) competitions. - Result: **all frontier models given $100K lost money** — captured public imagination that "AI is not as great as they thought." ### Why models fail at long horizons (his diagnosis) - Industry is **too biased toward coding/procedural tasks** ("do this, fix that" — limited solution space, little creativity). - Not enough **open-ended tasks**; real-world complexity and uncertainty not captured by benchmarks. - Insufficient **simulation of other players/agents** with different goals. ### Compute trade-offs - Must balance GPU use between training and inference; **pipeline RL** trains while sequences are still generating, trading off-policy-ness for GPU utilization. - From experience: **up to ~8 off-policy steps is okay**. - Problem: long-horizon inference may take weeks, exceeding the 8-step constraint — GPUs idle. - **Bootstrapping via value model** (generate expected value before episode end, "like dopamine in the human brain") allows continued training — but introduces **value model bias**. Trade-offs persist. ### Infrastructure - **Open Review AI** (openreview.ai): GR platform hosting 350+ environments via a single API endpoint, used internally and by some frontier and new labs. ### Summary position - Excitement lies in long horizons, requiring new thinking on **algorithms, environments, and compute**; many challenges and trade-offs, but long horizon is "a mindset" and GR's mission. --- ## Preserved Tensions & Qualifications - Ross's contradiction-with-outcome framing: Galactica was "a bloody good model" (SOTA on scientific benchmarks) yet failed publicly — reinforcing his RL-necessity claim. - His SOTA internal reasoning recipe produced results *without* the emergent reflection of R1/O1 — an unresolved "why" he answers only retrospectively (base model quality). - Chengxi repeatedly flags **trade-offs** rather than solutions: off-policy vs. GPU utilization; bootstrapping vs. value model bias; archives vs. cheating risk. - His diagnosis of industry benchmark bias ("I believe now...") is presented as his personal assessment, not established fact. # Video Attribution Record ## Video Metadata - **Title:** Scaling to Long Horizons — Ross Taylor & Chengxi Taylor, General Reasoning - **Video ID:** `01a15ca8-092a-70fe-99d7-931887d4117c` - **Format:** Two-part conference talk with speaker handover ## Speaker Roles ### Ross Taylor - **Role:** Creator/co-presenter (first segment) - **Self-identified credentials:** - CEO of GR ("London-based reinforcement learning company") - Former reasoning lead at Meta AI (worked on Llama models, Galactica) - Former member of the Papers With Code team (acquired by Meta in 2019) - **Segment scope:** Personal history of language modeling (~2020–2023), Galactica, RLHF, Llama post-training, unpublished "RSP" recipe ### Chengxi Taylor - **Role:** Guest/co-presenter (second segment, introduced by Ross) - **Self-identified credentials:** - Co-founder and president of General Intelligence Inc. (GR / General Reasoning) - **Segment scope:** Long-horizon RL challenges — context compaction, value models/critics, benchmarks, compute trade-offs, infrastructure ## Referenced Sources & Third-Party Claims - **Models/projects cited by Ross:** Galactica, ChatGPT (GPT-3.5), InstructGPT (2022), Llama 2/3, PaLM (540B), Chinchilla, GPT-4, OpenAI O1, DeepSeek 01/R1 (referred to as "DeepMind"/"Deep Seek" in transcript), RLHF pipeline - **Claims attributed to Ross's firsthand experience:** - Unpublished internal Meta "RSP" recipe (continued pre-training + PPO with verifiable rewards + value model initialization) - Galactica benchmark figures (68% vs GPT-3.5's 49%; chain-of-thought 36% vs PaLM's 19%) - **Claims attributed to Chengxi:** - Kelly bench benchmark (General Reasoning); all frontier models given $100K lost money trading on Premier League football matches - Coverage claim: featured on the front page of the *Financial Times* - Empirical claim: off-policy training "up to eight steps is okay" (stated as team experience) ## Promotional Segments (Flagged) - **Chengxi — Open Review product promotion:** - "Open Review was a product is a platform by General Reasoning... check it out. openreview.ai" - Claims: 350+ environments, single API endpoint, used internally and by "some frontier labs and new labs" - **Chengxi — closing company promotion:** - "If you are interested, follow us. General Reasoning, we're London-based AI research company" - **Ross — soft promotion** of second segment: "a really talk with a lot of alpha, so I'd encourage you to stick around" ## Availability & Limitations - **Comments:** Not addressed in transcript; no curated metadata supplied — **unknown** - **Curated metadata:** None supplied beyond the transcript itself - **Transcript quality note:** Contains apparent transcription errors (e.g., "Deep Seek 01" vs DeepSeek; "General Intelligence Inc." vs "General Reasoning" naming inconsistency; "Kelly bench" name uncertain from transcript alone) ## Uncertainty Notes - Company naming is inconsistent in the transcript: Ross says "GR"; Chengxi says both "General Intelligence Inc." and "general reasoning" — exact legal/entity name is **not confirmed** by this source. - "RSP" recipe details are reported as **unpublished** work; no external citation exists in the transcript. Uh so this talk is called scaling to long horizons. My name is Ross. Uh I'm the CEO of GR. We're a London-based reinforcement learning company. Uh before GR, I was the reasoning lead at Meta AI working on Llamas, uh Galactica, lots of other models back in the day. I'm joined by Chengxi, uh co-founder and president of GR. Uh and yeah, hit today we're going to talk about algorithms, environments, compute, all the things you need to do to get agents scaling uh to kind of long with tasks. So we're going to have two parts of this talk today. I'm going to first of all start with a personal perspective about, you know, the early days, the golden age of language modeling in between like maybe 2020 and 2023. Uh I'll talk about, like I said, all these models and some of our early reinforcement learning efforts for LLMs. And then Chengxi is going to talk about, you know, what's ahead, you know, what the next frontiers. And yeah, that's going to be a really talk with a lot of alpha, so I'd encourage you to stick around for that. So the journey so far. So my journey started here. Um so this was the Papers With Code team. I'm sure many of you used Papers With Code back in the day. So we were a London-based startup 2019. Uh we were acquired by Meta later that year. And then we had a crazy transition within Meta to do research. Um so we did, like I said, Galactica. Then after ChatGPT came out, we started the post-training for Llama 2, Llama 3. So all the great work you uh saw there was folks in this room. And lots of other interesting stuff that never got published as well. Uh reasoning Llama and lots of other things. Um so yeah, this this small team I'd like to think like the open weight kind of revolution started in this room. And you know, it really like hit home this idea to me that kind of small focused teams, like even in like the age of scaling, can do amazing things if people are aligned. Now for me, uh things got particularly crazy in 2022. Uh so let me tell you a story. Um the media perception is that ChatGPT came out of nowhere, you know, shocked the world, and that's how kind of the modern AI wave started. But, you know, I have a different personal perspective on this because 2 weeks before ChatGPT came along, there was another language model called Galactica. So let's talk about Galactica. Galactica and, you know, ChatGPT, you know, they were both, you know, in some respects quite similar. They're both based on pretty good base models. Galactica itself was a base model, and then ChatGPT was based on GPT-3.5. But there was a clear difference in outcomes. So Galactica at the time shipped with this slight base model demo. And as you guys know now, like base models, they come with a lot of quirks. You know, they hallucinate. You prompt them to do, you know, silly things, they would do silly things. Whereas ChatGPT wasn't just a base model, but had this like crucial reinforcement learning from human feedback pipeline. And this was the key thing that made LLMs like really products for the first time. So I like to think in a weird kind of way, this is like the first like kind of natural experiment showing you that kind of RL like provides value, right? Uh and to my misfortune, it was like a very personal like kind of a natural experiment and you know, Galactica blew up. Uh but that's like a good like kind of lesson there. A good base model is not enough. So I took that lesson quite early on. So like I said, RLHF made LLMs products. They were the thing that kind of made LLMs cross the Rubicon into something that wasn't just a toy, but used by now billions of people. But you didn't have to wait until ChatGPT to see this. Like even at the time, like InstructGPT in 2022 had these pretty stunning results. Like a 1 billion parameter model with RLHF was outperforming 175 billion models. So two orders of magnitude fewer parameters, but getting better results. So that was astonishing. So if you were paying attention closely, you know, maybe we should have been as well, but we were focused on a, you know, bloody base model, which is hard work in 2022. But that shows you how you know important even basic RL is. And the Galactica demo itself, I mean, it set off a storm. So, ancient history now, but we put out a demo. We let people play around with it um because it's kind of cool. And at the time people got scared. So, it was like, you know, like I said, you prompt it on like a research paper on Dyson spheres or a report on the benefits of eating crushed glass and people are like, "Oh my god." Um so, that was the state of things in 2022. And yeah, I'll be honest, the meta association didn't you know help us either. Um and yeah, the tragic story in a way was a lot of the, you know, novel work was maybe overshadowed. But you know, the paradox of this whole thing is that at the time Galactica was actually a bloody good model. Um like it outperformed Palm, Chinchilla, GPT-3.5 with a lot less compute in scientific domains. It was state-of-the-art. So again, that also shows you how powerful RL is. You can have a sota base model, but that is not enough. Um so, here you see on like a math who's kind of beating Chinchilla. Uh kind of latex equations, you know, science, you know, it was getting around 68% compared to GPT-3.5, 49%. So, crushing there. And chain of thought as well. So, you know, a Palm at the time which was a Google Brain model, 540 billion. You know, that's 30 billion in Galactica was getting like 36 versus 19%. So, double the performance, order of magnitude less results. So, that again reinforces base good base models not enough. But it introduced some key ideas which I think are very important. I mean, Galactica was the first LM to really crack data efficiency. 105 billion uh token corpus compared to you know, trillion uh tokens in Chinchilla. And it was a really contrarian at the time cuz you know, at the time everyone was like, "Okay, we just need more tokens." And Galactica said, "No. High quality, you know, curated data sets really matter." And that was a real driver of those results you just saw. And it was also like the first major LM to really crack multi-epoch training. It sounds ridiculous now, but at the time the consensus was you don't do more than an epoch. Um but this kind of rule of thumb, you may have heard of it like four uh epochs repeated data, that was formalized later. but Galactica was the first like real empirical result for that. Now, perhaps more importantly, there's this idea of thinking tokens. And some of you might remember this, but it was like really quite buried within the paper. So, around this time there were like different ideas for reasoning. There was chain of thought, which was one idea. It's where you prompt for like kind of like the steps. There were scratch pads where you just like put like very like numerical kind of intermediate steps. But, Galactica was really this first idea which said, "No, this is an internal working memory process. This is an internal thinking. You should be inside these tags, and you should spend the inference computer before you get to an answer, right?" So, these are all like quite like pressing ideas. But, you know, Galactica came and went, blew up, and then, you know, we were kind of tasked as a team to kind of spin up the post-training effort for Llama. But, I had like a personal obsession, which was like reasoning. And I had like a really simple idea at the time, which is, "What if we applied kind of reinforcement learning pressure to this like kind of thinking tags as work?" Like, what if we just optimize the thing in between the thinking? And if that sounds familiar, then this is like kind of what Deep Seek 01 ended up doing 2 years later. But, there's a key difference. So, at the time we only had Llama 2 base models, terrible mathematics corpus, terrible results on math. And, you know, the context window, you know, we're all like, you know, very context-rich now, 1 million tokens. You know, the back in the day it was 4,000. It wasn't too fun. But, we still had a recipe at the time, and this is unpublished, but it was a really good for the meta. So, RSP was this. Number one, continue pre-training of Llama 2 towards mathematics and science data. So, that's the first thing. Llama 2, math corpus, so let's fix that. Number two, PPO with verifiable rewards. But, notice this isn't GRPO, right? So, we had a time and a strong outcome reward model to initialize the value model. That was a key thing, lots of data on value models at the time. And internally at the time, this kind of recipe led to state-of-the-art results on math and reasoning. So, we were kind of like, "Wow, this is like really shows the power of having the right objective." But, the really fascinating thing is like we had great results, but we didn't have like inference time scaling. We didn't have this reflective behavior that became like the hallmark of R1 and O1. You know, back weights, you know, back tracking, all this kind of stuff. So, it begs like the question like, "Why? Well, why didn't we have that moment?" And we got an answer around like 2 years later. So, there's a couple of like things going on here, but essentially better base models were the thing that really got RL cooking. And when DeepMind came out, I was kind of shocked at the time. I was like, "Holy we just like tried the same thing. We didn't have this. What's going on here?" And in a weird kind of way, the real lesson was it was just like the bitter lesson, like the most purest form of bitter lesson possible. Like, better base models, more RL computes, bigger context windows, and that's all you need for this kind of emergent behavior. It also like says something like quite important about the sociology of like research because the fact that OpenAI had this model, you know, GPT-4 level model before anyone else, it allowed them to see further, right? So, that's a really interesting point. Like, the the age of scaling means that if you have certain prerequisites in place, you become smarter, you see further, you see more ideas. So, really interesting point. So, this was me in 2024. I was a very sad panda, uh defeated by ChatGPT and O1. But, I wasn't deterred. Um so, I wanted to seek the next wave, and I still was like convinced that kind of reasoning hadn't been solved. So, we started GR uh to take on truly like big tasks. And with that in mind, I'm going to hand over to Chengxi. He's going to talk about what we're kind of thinking about now. \>\> Ross. Ross. Ross. \>\> Hi, everyone. I'm Chengxi Taylor, co-founder and president of General Intelligence Inc. I'm going to share what it takes to scale to long horizon. First, I want to make it clear. Long horizon task is not just an engineering problem. It is a mindset. If we want to solve humanity's biggest problems, such as cure cancer, solve millennium's prize problem, or go into Mars, we have to be patient. It take time. And if we want AI to move us towards that level impact, we have to think about long horizon. But here's the first problem. We have a scarce context window. If you take Fermat's Last Theorem as example, what it take for the mathematician was over 10 years time of reading paper, writing down thoughts in the scratch pad, or taking a walk to generate the creative ideas. If we convert to token, that's probably tens of a billions of even hundreds billions. But where we are now, just a 1 million token context window. So one solution is use compaction. So what it essentially does is generate the token until the end of the context window, summarize, and then on top of that generate more tokens. And the beauty of applying RL in the situation is kind of like kill two birds with one stone. You apply RL to the compaction and also the task. But here's the problem. With long horizon, there are three issues. The first is the gradient variance scales with the length. And the second is a sparse reward. And you have this credit assignment problem. And finally, there's also variable length of the trajectory that adds to the problem of optimization. So to solve this issue, we can apply critics, which is the value model. And value model can reduce the variance and also have a couple advantages, such as on the trajectory level that fits compaction very well, and also encourage the batch diversity. And also, I'll talk later on bootstrapping. Basically, get signal before the end of the episode. But, the downside for this is that um it's more complicated than GRPO, and basically, you have to train another value model alongside with the policy model. And there's some tools to help with the context limitations, such as a file system tools, which essentially like a scratchpad for AI to write to the uh reasoning thought. And self-search tools, which allows agent to search over the previous trajectory. And then you have archive tools. In the case like auto research, you can build upon your previous result. But, we have to be careful. In other scenario, you don't want an AI to cheat by just to grab the previous answer without thinking. So, how good are the current model on this long horizon task? In the general reasoning, we construct the benchmark called a Kelly bench, where we're actually featured in the front page of the Financial Times. Kind of a caught us off guard how much the mainstream have interest in this. So, basically, what we did is that we allowed the agents to build machine learning models to um trade in the football matches over 1-year horizon. In this case, it's a Premier League, if you're interested in football. And we're so fascinated by this because there's a real money to be made, and if it was a successful, couldn't make a billions. There's a whole industry on sports betting. And unlike things like a cargo competition, this has a real-world implication. But, here's the result. As you can see, we gave all the frontier models a 100K to start. All of them lost her Sad. And that captured the public's imagination. Oh, AI is not as great as they thought. And why are models so bad at long horizon? First, I believe now the AI industry is a little bit too biased towards coding and procedure task. What I mean is that the current task is a most formulated like do this and fix that. Normally that limits the solution like one or two. There isn't just too much space for creativity. And second of all, not enough focus on open-ended task. We live in the real world with a lot of a complexity, uncertainty, and that's not fully captured by the current benchmark. And also, there isn't enough simulation of the real world. We live in the world that there are other players. Like in today's conference room, there are other real people who have a different thought, a different games than you have in your mind. That's the complexity that's not fully captured. And another thing I want to talk about is the long horizon impact on compute. We know that GPUs are scarce and precious resources. And in this case of a long horizon reasoning, you have to be careful about how to optimize your use between training and inference. And pipeline RL is a quite popular technique nowadays. So basically, it's a trade-off between off-policy and the GPU utilization. So traditionally, you let inference run towards the end and then you start to train the model. But in the case of long horizon, you have to wait until the inference finish. What the pipeline RL does is that you let the sequence to be generated and you start to train the model while there's a still more sequences being generated. And you see this created off-policy. But from the our experience, normally off-policy up to eight steps is okay. So, essentially, we made a trade-off between the off-policy and the GPU utilization. But, here comes the issue. As we the long horizon indicates, sometimes the inference would take weeks or even more. In that case, inevitably, it will goes beyond the constraint of the eight steps of our policy. So, your GPU have just to sit there idle and wait for it to finish. And if you don't want to wait, as I mentioned before, applying the value model allows you to bootstrap. What it means is that before the end of the episode, you generate expectation. It's like a dopamine in human brain. And that allows you to train the model. But, here's another trade-off. While you utilize the GPU fully, you introduce the value model bias. So, there's always a bit of trade-off in those solutions. And I want to also mention that in the long horizon, infrastructure is important, especially for the environment. And Open Review was a product is a platform by General Reasoning. If you're interested, you can check it out. openreview.ai. So, it's a place where host over 350 environments and with a single API endpoint. And we use this for our internal RL and also some frontier labs and new labs are using this. So, to summarize both Ross and my speech, it has been a long journey as long horizon indicate. We as a team have seen the paradigms in AI reasoning on pre-training and agents in the past few years. But, looking ahead, what makes us really excited is the long horizon. And it requires us to think, have a new thinking on the algorithm, environments, and compute. There are a lot of challenges and trade-offs, but we find it's really exciting to take on this journey because, as I mentioned in the very beginning, long horizon is not just engineering problem. It is a mindset, and if we really are ambitious to solve humanity's biggest problems, this is the journey for everyone. And that's also the mission for general reasoning. So, if you are interested, follow us. General Reasoning, we're London-based AI research company. Thank you. \>\> \[music\]