Interconnects

• GPT-4.5: "Not a frontier model"?

  More: https://www.interconnects.ai/p/gpt-45-not-a-frontier-modelAs GPT-4.5 was being released, the first material the public got access to was OpenAI’s system card for the model that details some capability evaluations and mostly safety estimates. Before the live stream and official blog post, we knew things were going to be weird because of this line:GPT-4.5 is not a frontier model.The updated system card in the launch blog post does not have this. Here’s the original system card if you need a reference:Regardless, someone at OpenAI felt the need to put that in. The peculiarity here summarizes a lot of the release. Some questions are still really not answered, like “Why did OpenAI release this?” That game theory is not in my purview.The main contradiction to the claims that it isn’t a frontier model is that this is the biggest model the general public has ever gotten to test. Scaling to this size of model did NOT make a clear jump in capabilities we are measuring. To summarize the arc of history, the jump from GPT-3.5 to GPT-4 made the experience with the models go from okay to good. The jump from GPT-4o (where we are now) to GPT-4.5 made the models go from great to really great.Feeling out the differences in the latest models is so hard that many who are deeply invested and excited by AI’s progress are just as likely to lie to themselves about the model being better as they are to perceive real, substantive improvements. In this vein, I almost feel like I need to issue a mea culpa. I expected this round of scaling’s impacts to still be obvious before the brutal economic trade-offs of scaling kicked in.While we got this model, Anthropic has also unintentionally confirmed that their next models will be trained on an approximation of “10X the compute,” via a correction on Ethan Mollick’s post on Claude 3.7.Note: After publishing this piece, I was contacted by Anthropic who told me that Sonnet 3.7 would not be considered a 10^26 FLOP model and cost a few tens of millions of dollars to train, though future models will be much bigger.GPT-4.5 is a point on the graph that scaling is still coming, but trying to make sense of it in a day-by-day transition is hard. In many ways, zooming out, GPT-4.5 will be referred to in the same breath as o1, o3, and R1, where it was clear that scaling pretraining alone was not going to give us the same level of breakthroughs. Now we really know what Ilya saw.All of this marks GPT-4.5 as an important moment in time for AI to round out other stories we’ve been seeing. GPT-4.5 likely finished training a long time ago — highlighted by how it has a date cutoff in 2023 still — and OpenAI has been using it internally to help train other models, but didn’t see much of a need to release it publicly.What GPT-4.5 is good forIn the following, I am going to make some estimates on the parameter counts of GPT-4.5 and GPT-4o. These are not based on any leaked information and should be taken with big error bars, but they are very useful for context.GPT-4.5 is a very big model. I’d bet it is well bigger than Grok 3. We have seen this story before. For example, GPT-4 was roughly known to be a very big mixture of experts model with over 1T parameters total and ~200B active parameters. Since then, rumors have placed the active parameters of models like GPT-4o or Gemini Pro at as low as 60B parameters. This type of reduction, along with infrastructure improvements, accounts for massive improvements in speed and price.Estimates place GPT-4.5 as about an order of magnitude more compute than GPT-4. These are not based on any released numbers, but given a combination of a bigger dataset and parameters (5X parameters + 2X dataset size = 10X compute), the model could be in in the ballpark of 5-7T parameters total, which if it had a similar sparsity factor to GPT-4 would be ~600B active parameters.With all of these new parameters, actually seeing performance improvements is hard. This is where things got very odd. The two “capabilities” OpenAI highlighted in the release are:* Reduced hallucinations.* Improved emotional intelligence.Both of these have value but are hard to vibe test.For example, SimpleQA is a benchmark we at Ai2 are excited to add to our post-training evaluation suite to improve world knowledge of our models. OpenAI made and released this evaluation publicly. GPT-4.5 makes huge improvements here.In another one of OpenAI’s evaluations, PersonQA, which is questions regarding individuals, the model is also state of the art.And finally, also GPQA, the Google-proof knowledge evaluation that reasoning models actually excel at.At the time of release, many prominent AI figures online were touting how GPT-4.5 is much nicer to use and better at writing. These takes should be taken in the context of your own testing. It’s not that simple. GPT-4.5 is also being measured as middle of the pack in most code and technical evaluations relative to Claude 3.7, R1, and the likes.For an example on the writing and style side, Karpathy ran some polls comparing GPT-4.5’s writing to GPT-4o-latest, and most people preferred the smaller, older model. Given what we know about post-training and the prevalence of distilling from the most powerful model you have access to, it is likely that GPT-4o-latest is distilled from this new model, previously called Orion, and its drastically smaller size gives it a night and day difference on iteration speed, allowing for better post-training.More on the character in that GPT-4o-latest model was covered in our previous post on character training.All of this is a big price to pay to help OpenAI reclaim their top spot on ChatBotArena — I expect GPT 4.5 to do this, but the results are not out yet.I’ve been using GPT-4.5 in preparation for this. It took a second to get used to the slower speed, but it’s fine. I will keep using it for reliability, but it’s not worth paying more for. o1 Pro and the other paid offerings from OpenAI offer far more value than GPT-4.5.Interconnects is a reader-supported publication. Consider becoming a subscriber.Making sense of GPT-4.5’s ridiculous priceWhen the original GPT-4 first launched, it was extremely expensive. In fact, GPT-4 was comparable in price to GPT-4.5 at launch. Here’s a help post on the OpenAI forums, conveniently found by OpenAI DeepResearch with GPT-4.5, that captures the context. GPT-4 launched in March 2023.We are excited to announce GPT-4 has a new pricing model, in which we have reduced the price of the prompt tokens.For our models with 128k context lengths (e.g. gpt-4-turbo), the price is:* $10.00 / 1 million prompt tokens (or $0.01 / 1K prompt tokens)* $30.00 / 1 million sampled tokens (or $0.03 / 1K sampled tokens)For our models with 8k context lengths (e.g. gpt-4 and gpt-4-0314), the price is:* $30.00 / 1 million prompt token (or $0.03 / 1K prompt tokens)* $60.00 / 1 million sampled tokens (or $0.06 / 1K sampled tokens)For our models with 32k context lengths (e.g. gpt-4-32k and gpt-4-32k-0314), the price is:* $60.00 / 1 million prompt tokens (or $0.06 / 1K prompt tokens)* $120.00 / 1 million sampled tokens (or $0.12 / 1K sampled tokens)GPT-4.5’s pricing launched at:Input:$75.00 / 1M tokensCached input:$37.50 / 1M tokensOutput:$150.00 / 1M tokensOpenAI included language in the release that they may not keep this model in the API, likely forecasting low demand, as they wanted to hear from users if it enabled entirely new use-cases.Many analysts think that Nvidia’s next generation of GPU, Blackwell, which comes with GPUs with far more memory per FLOP (enabling storing bigger models), are not priced into this. We can expect to see the same arc of pricing with 4.5 as we did with 4 to 4 Turbo to 4o.* GPT-4 Turbo launched in November 2023 at $10 / 1M input and $30 / 1M output.* GPT-4o launched in May 2024 at $2.5 / 1M input and $10 / 1M output.These are huge reductions, about 10X.These are products that OpenAI makes a healthy margin on, and there are no signs that that isn’t the case. The AI community collectively has grown so accustomed to incredible progress in making the technology more efficient that even a blip in the process, where bigger models are available, feels potentially bubble-popping.The future of scalingScaling language models is not dead. Still, reflecting on why this release felt so weird is crucial to staying sane in the arc of AI’s progress. We’ve entered the era where trade-offs among different types of scaling are real.If forced to summarize all of this curtly, it would be: GPT-4.5 is, oddly, ahead of its time.This means that the progression of AI needs to take a different tack, but we already knew this with the rapid progress of reasoning models. The true impact of GPT-4.5 is when it is integrated with multiple lines of rapid progress.One of the flagship results in the DeepSeek R1 paper and related RL follow-up work in the AI community is that scaling RL training works better on bigger models. There is a lot of work to do to know all the domains that’ll be absorbed into this umbrella. Future models like o4 could be distilled from a reasoning model trained on GPT-4.5. In fact, this may already be the case. OpenAI’s current models likely would not be so good without GPT-4.5 existing.In as soon as a year, most of the models we are working with will be GPT-4.5 scale and they will be fast. The “well-rounded” improvements they offer are going to help make many more applications more robust, but OpenAI and others in the AI labs have pushed scaling a bit further than the current serving infrastructure can support.Frontier labs are not taking enough risk if they’re not going to try to push the limits of every direction of scaling they have. Though releasing the model isn’t needed, we have to guess why OpenAI actually wanted to do this. It’s likely that GPT-4.5 is being used in other internal systems for now and other external products soon, so releasing it is a natural step on the way to the next thing, rather than a detour.GPT-4.5 is a frontier model, but its release is not an exciting one. AI progress isn’t free, and it takes a lot of hard work. Most people should only care when GPT-4.5 is integrated into more than just chat. Get full access to Interconnects at www.interconnects.ai/subscribe

  --------  

  10:02
• Character training: Understanding and crafting a language model's personality

  https://www.interconnects.ai/p/character-trainingThe vast majority of evaluations used to measure progress on post-training at frontier laboratories are internal evaluations rather than the evaluations you hear about all the time like MATH or GPQA. These, the well-known intra-industry evaluations, are certainly important for ballparking behavior, but for every public evaluation, these frontier laboratories are likely to have 10+ fine-grained internal evaluations.The internal evaluations these model providers have cover a range of topics. Surely, most of them are basic, repetitive user behaviors that they need to make sure a new model doesn’t roll back too many of. Of these, the vast majority are likely skills, and “character” remains more of an art than a hill to climb up with careful data engineering.Leading post-training laboratories surely know how to reinforce more robust behavior within a specific character, as seen by the march of progress on evaluations like ChatBotArena, but crafting a specific personality from scratch is an open question.The primary goal of this post is to start the conversation outside of frontier AI labs around character training. Character training is the subset of post-training designed around crafting traits within the model in the manner of its response, rather than the content. Character training, while being important to the user experience within language model chatbots, is effectively non-existent on the web.We don’t know the trade-offs of what character training does, we don’t know how exactly to study it, we don’t know how much it can improve user preferences on ChatBotArena, and we should.The appearance of the AIs people are using is deeply coupled with how intelligent users will find it to be. Style of communication is crucial to how information is parsed. This is likely a very high priority to industrial labs, but something that almost no academic literature exists on. Even though I want to do research on this, I’m honestly not sure how to do so yet other than a 1 of 1 technical report on findings.ChatGPT gets character depthOut of nowhere on Saturday, February 15th, Sam Altman tweeted about this new GPT-4o model that will serve as the foundation of ChatGPT.This is the biggest subjective change I’ve ever felt within intermediate model versions, from any primary provider — something more akin in vibes change to the shift from GPT-3.5 to GPT-4. The model immediately and consistently showed new behavior patterns. I found these very positive (Karpathy agrees), but they’ll take some getting used to.Where ChatGPT used to sound robotic and shallow, it’s now very clearly leaning into a chipper assistant demeanor. Yes, for basic tasks, this new default model in ChatGPT is very Claude 3.5-like — more testing is needed to know if this GPT-4o with its peer models like o3-mini can dethrone Claude 3.7 Sonnet as a daily programming driver.The biggest changes in the new GPT-4o model are:* It now loves to reference past interactions in the chat (way more obvious than any other provider has been) — it was trying to flex that it knows my dog breed, mini schnauzer, or my book topic, RLHF. This is very in line with the new roadmap to GPT-4.5 and GPT-5 that Altman posted, where ChatGPT is designed around a fluid experience rather than standalone, siloed, powerful models.* The model is very chipper, sprinkles in more emojis, and is almost funny.* The multi-turn conversation is more dynamic, with follow-up questions and added texture to longer back and forths.The reasons are at a high level very complementary to those I listed when I switched to Claude as my daily driver model.The shocking part of this is that the impact of this sweeping change is almost entirely undocumented. Yes, OpenAI updated the Model Spec (my previous coverage here and here), but that doesn’t really capture how this model is different — it just clarifies the direction OpenAI is optimizing for. There are a few overlapping interpretations of this lack of transparency:* OpenAI cannot precisely measure the differences as a few specific behavior traits, so they can only see the model performs better in high-level testing like ChatBotArena or other A/B testing, but they cannot capture the changes in score deltas between a few evaluations they could release.* AI is moving so fast that taking the time to document these models is not worth it,* Detailing the changes will make the character too easy to reproduce and will be another path of “distillation” of OpenAI’s models.The community of model users is extremely far from having clear ways to measure these differences. While there are vibe tests on Twitter, they will not be conclusive. ChatBotArena won’t even come close to measuring the levels of these differences (and in the case of referencing past chats, it cannot). Character training is the sort of addition to a post-training stack that takes industrial training techniques from being reproducible, but expensive, to dark arts that are largely undocumented.The most interesting part of the model spec for industry analysts is this plot where OpenAI shares the agreement rate of their newer models. This is comparing a reasoning model, o1, to a GPT-4o model, so there are questions of whether this is attributable to reasoning training.Every frontier AI laboratory should have a model specModel Specs are the sort of community norm that a race to the top is the goal. They’re muddled if mandated — how would you actually check that a required model spec is accurate? — but if they are implemented by every lab carefully with feedback from the community, it would be far easier for the development ecosystem to exist around models.The model spec is an extremely useful document detailing how developers can expect your models to change over time. They are also one of the few sources of insight we have into what the model providers are trying to get their models to do (which has regulatory advantages) and let us know what is an intentional or unintentional behavior mode.Interconnects is a reader-supported publication. Consider becoming a subscriber.A model spec doesn’t provide all the information we need to keep up with model versions. This new version of ChatGPT desperately needs to be accompanied by evaluations capturing the behavior change, otherwise, a lot of undocumented differences will be passed on to developers updating endpoints to it. This is another rendition of the same lack of transparency we’re used to from leading AI laboratories.The closest thing Anthropic has to a model spec is the mix of Claude’s Constitution and this blog post on Claude’s Character. Character training is a fairly new technique for the industry. From Anthropic’s post:Claude 3 was the first model where we added "character training" to our alignment finetuning process: the part of training that occurs after initial model training, and the part that turns it from a predictive text model into an AI assistant. The goal of character training is to make Claude begin to have more nuanced, richer traits like curiosity, open-mindedness, and thoughtfulness.The process is extremely synthetic data-heavy, but requires an artist’s touch, as stated later in the blog post: It “[relies] on human researchers closely checking how each trait changes the model’s behavior.”Character training being the focus of developments is the strongest endorsement that RLHF and related approaches have shifted from their philosophical motivations of alignment to being primarily an empirical tool. The models can capture so many different behaviors, but getting them to reliably behave how we want is the hardest part. Right now, it seems more likely that this is about capturing the upside of RLHF as a performance tool, rather than a safety one.One of the few public discussions of character training came from Amanda Askell during her appearance on the Lex Fridman Podcast (taken from the transcript):Lex Fridman (03:41:56) When you say character training, what’s incorporated into character training? Is that RLHF or what are we talking about?Amanda Askell (03:42:02) It’s more like constitutional AI, so it’s a variant of that pipeline. I worked through constructing character traits that the model should have. They can be shorter traits or they can be richer descriptions. And then you get the model to generate queries that humans might give it that are relevant to that trait. Then it generates the responses and then it ranks the responses based on the character traits. In that way, after the generation of the queries, it’s very much similar to constitutional AI, it has some differences. I quite like it, because it’s like Claude’s training in its own character, because it doesn’t have any… It’s like constitutional AI, but it’s without any human data.In summary, Anthropic uses the same techniques they use for Constitutional AI and general post-training for capabilities to train these models’ characters. This is not surprising. This could be related to Askell’s other Tweet on how she designs system prompts, as system prompts are the easiest way to quickly change a model’s character:The boring yet crucial secret behind good system prompts is test-driven development. You don't write down a system prompt and find ways to test it. You write down tests and find a system prompt that passes them.This is very in line with what we started this post on — internal AI lab evaluations.How far can you push character training?Ahead of the Grok 3 release, Elon Musk Tweeted this example from Grok 3, saying it was “based.”One of the predominant reactions to Grok 3 was, “Wait, so it isn’t actually based?” This is one of the big questions of character training and lacking model specs. Did xAI not figure out how to make their model-based and reliable? What model was Elon using here?Whatever your politics, it’s likely that the default personality of models that you encounter will eventually not be something you like. There’s quite a lot of nuance in what the perfect chatbot is for each user.Companies should be allowed to have a default personality for the models of their choosing, but a far better long-term equilibrium is to make the expectation that model providers make it easy to get exactly the personality you like out of a model. This isn’t regulation I’m recommending right now, but one way to make sure that an all-powerful AI model isn’t going to reinforce one point of view is to have tests that models need to pass on the breadth of their character and views.Model specs are a step in the right direction to avoid drama about “what did they actually want their model to say,” but we still have a lot of work to do on creating a spectrum of tools that captures all the relevant information when comparing models. Get full access to Interconnects at www.interconnects.ai/subscribe

  --------  

  11:39
• Claude 3.7 thonks and what's next for inference-time scaling

  On Monday, February 24th, 2025, Anthropic announced their latest model, Claude 3.7 Sonnet, which is their first model explicitly trained to use more inference time tokens to improve performance. This is another reinforcement learning (RL) trained model (mentioned in system card). With this model, they also released Claude Code as a limited research preview, which is a “command line tool for agentic coding.” Continuous improvements in models are enabling new modalities and domains addressable by the models, but assessing the impact of each new domain takes far more time than a quick model reaction post.This is a tidy release, a solid improvement, but not a step change for Claude or the industry. Expect a lot of small changes to accumulate massively this year.Claude 3.7 Sonnet is a clear improvement over Claude 3.5 Sonnet (New) and continues to push the limits in areas where users love Claude (e.g. read Ethan Mollick’s review here). The scores for those areas such as software development (SWE Bench) and tool use, are clearly state-of-the-art. For example, Claude 3.7 Sonnet is the highest scoring “standard non-reasoning” language model on the Aider Polyglot benchmark. While models like o3 and Grok 3 DeepThink highlight superhuman performance on code benchmarks, this sort of behavior being integrated without extra inference time compute is wonderful. The price for superhuman coding AI is plummeting.Even with o1 Pro, I still find myself using Claude 3.5 (New) on a regular basis. O1 Pro is the best model for doing succinct, one-off tasks like writing short scripts. It is extremely controllable and will often work out of the box. Though, when I’m doing tricky, iterative tasks I still use Claude. Claude 3.7 Sonnet only makes these workflows stronger and I’m stoked to play with it further.The most useful piece of this release for those trying to understand the direction of the ecosystem, rather than just the status of today, is Anthropic’s post on Claude’s extending thinking where they detail the product trade-offs, alignment, and future of inference time compute in their models. Anthropic’s offering of extending thinking to boost inference-time performance is far, far cleaner than that of OpenAI’s current model drop down disaster. Anthropic’s thinking model is the same as their general purpose model, much like xAI’s Grok 3, and what OpenAI teased will be the plan for GPT-5. Having just one model makes lots of infrastructure, product, and training decisions cleaner, but may come at the cost of the absolute Pareto front of performance for your organization shrinking. The reasoning training being embedded in one model with a standard inference mode will make the reasoning benefits and behavior feel closer to something like Gemini-Thinking, rather than OpenAI’s o1 or DeepSeek R1 that are designed solely for this reasoning mode of operation. It doesn’t mean that in the limit that a single model will be weaker in performance, but rather that currently training them may be slower to iterate on than a “full” reasoning language model.Focusing on deploying just one model that serves all the users is one of many examples where leading AI companies are needing to make their offerings legible to users and easy to use — a sign of the industry maturing from a race to intelligence to a race to usefulness.Still, Claude’s interface is not perfect by any means, the user still has to intentionally go to a drop down menu to get performance when they need it. The best mode is that the model knows when inference compute is needed on its own. My hypothesis is that when training one model with reasoning and without, having the model figure out how much compute to use is harder than a reasoning-only model like o1 figuring out its own compute budget. Or, Anthropic needed to keep a special flag that is turned on and off in the system prompt. This is a subtle potential trade-off of putting reasoning in just one model, but we’ll see where the final equilibrium is.On the other hand, Claude 3.7 Sonnet is showing the reasoning traces directly to users like DeepSeek R1 and Grok 3. These organizations have different ways of saying why, but it is clear that users just enjoy seeing it and it builds trust. Anthropic, understandably is using the reasoning traces to monitor the alignment of the models. The reasoning chains in these models are how the general public is learning more about the internal representations of language models. Another interesting detail is that “didn’t perform our standard character training on the model’s thought process.” This is how Claude thinks out of the box and the actual answers have a different flavor to them. More research will study how far the reasoning chains can diverge from the answer language. We’ve seen research on latent reasoning within the model, but beyond this, we could have reasoning languages that are entirely ungrounded from human languages because they are a more token-efficient representation of information for the model. More on this soon.The developer facing version of this Claude Extending Thinking is far cleaner and a sign of things to come — developers can request a specific amount of thinking tokens in their response. How this works is that the model will stream thinking tokens until the number is reached, then shift to answer tokens. This is still one autoregressive stream and no search is being used in the new products, yet.This explicit control over the thinking and answering phases is a growing behavioral focus in training reasoning models — expect more here soon. Developers can tune the setting that works for them and keep it baked in, rather than relying on the user to pass in a query that just happens to get the model to think for a long time. Explicit test-time inference budget increases are much more covetable than needing to hit the gold mine in a prompt search.The best place to see where this could be applied is by selecting performance on a task that scales nicely with inference time compute. Anthropic ran a similar experiment on the challenging math evaluation AIME — the same one that OpenAI used in their original inference time compute plot. Here there’s a subtle difference from the developer experience, where in Anthropic’s internal tests the model could exit early. In practice, this subtle difference shouldn’t shift the usefulness of the deployment methodology.Anthropic continues in their excellent post, saying:Our researchers have also been experimenting with improving the model’s performance using parallel test-time compute. They do this by sampling multiple independent thought processes and selecting the best one without knowing the true answer ahead of time. One way to do this is with majority or consensus voting; selecting the answer that appears most commonly as the 'best' one. Another is using another language model (like a second copy of Claude) asked to check its work or a learned scoring function and pick what it thinks is best. Strategies like this (along with similar work) have been reported in the evaluation results of several other AI models).To accompany this, they shared the following results. It is crucial to note here that the dashed red line — pass@N — is not an actual evaluation result, but measuring if the correct solution appears in the number of answers generated on the X-axis. The two lines below show how good initial answer extraction methods are at selecting the right answer from the N candidates. As has been known for a long-time in inference-time scaling research is that the models can often generate the correct answer to extremely hard questions, but not reliably.Interconnects is a reader-supported publication. Consider becoming a subscriber.They make it very clear this is not used yet in their products:Parallel test-time compute scaling isn’t available in our newly-deployed model, but we're continuing to research these methods for the future.Still, this is a direction other labs are already pursuing. The best reporting on o1 Pro indicates that it does a “search” of some sort over parallel generations. Other OpenAI employees have stated that o3 uses a learned verifier to extract answers, at least for coding domains. As progress in scaling single-streams from the language model slows, this is the natural next place for scaling to turn to. As it has been for some time, performance limits are largely different forms of infrastructure problems before models can be served to users.Claude is here, and it reinforces that RL training is a short path to inference time scaling laws being used, but in the long-term we will have more methods for eliciting the inference-time tradeoffs we need for best performance.Thanks to Ross Taylor for some immediate feedback on this post. Get full access to Interconnects at www.interconnects.ai/subscribe

  --------  

  9:48
• Grok 3 and an accelerating AI roadmap

  Full post: https://www.interconnects.ai/p/grok-3-and-an-accelerating-ai-roadmapxAI launched their latest flagship model, Grok 3, last night via a live stream on X, which is a new take on the launch process, but it largely felt familiar. Grok 3 is a state-of-the-art model on some important benchmarks. The core is that it is state-of-the-art relative to available models and we know better models are out there. Only some of them have been announced, some of them have been teased, and others lie in waiting.What feels different is how the broader AI industry is signaling rapid progress coming soon. xAI said on the livestream that they will be updating the model “daily.” An era of sitting on unreleased models could be ending.Grok 3’s release is a reinforcement of trends people began reckoning with as of the release of DeepSeek V3 + R1 — AI progress is not held in the hands of a few companies nor is it slowing down. 2023 and 2024 were defined by truly state-of-the-art AI being concentrated within OpenAI, Anthropic, and Google, where these companies could take a lot of time to package models from training to release and still have a substantial moat on capabilities relative to their peers.At the time of R1’s launch, the “people’s choice” model was Claude 3.5 Sonnet, a model that had been trained “9-12 months ago” and the best models like Claude 3.5 Opus or GPT-4.5 (a.k.a Orion) were not available to users for a grab bag of reasons.Competitive pressure from DeepSeek and Grok integrated into a shifting political environment for AI — both domestic and international — will make the established leading labs ship sooner. A large portion of delays in delivering models is for “safety testing,” but we don’t have exact details on how much of it was that and how much was cost-benefit tradeoffs (and other big company hurdles such as legal departments). The brand, and culture, of “having the smartest model” is extremely important to these companies, but having a way smarter model was often financially too much to bear.“Safety” is actively being removed from the spotlight of the AI discourse. It is possible that this overcorrection causes meaningful harm, as this is an extremely powerful and rapidly evolving technology, but the political capital to make safety a core tenet of the AI industry was spent too early relative to meaningful harm emerging.Increased competition and decreased regulation make it likely that we, the users, will be given far more powerful AI on far faster timelines.We’ve seen time and time again the value of having the best model first. The only way to onboard new users is to have some capability or behavior that your model differentiates on. With the pace of progress high, minimizing the time from training to release is the best way to maximize one’s chance of impact.DeepSeek and xAI show how organizations with slightly trailing technical progress or resources can outshine the likes of OpenAI and Anthropic who have voluntarily not shipped their latest models.Interconnects is a reader-supported publication. Consider becoming a subscriber.Grok 3 by the numbersBenchmarks and vibe tests mark Grok 3 as one of the best models available today. As with any release, companies often choose evaluations that flatter their models. Yes, winning on these evaluations is extremely challenging, and much credit must be given to xAI for delivering a leading-edge model just about 19 months after its incorporation.That being said, what is shown below is a total of 4 language model evaluations. Given that models like DeepSeek R1 or Gemini Thinking launch with 10-20 evaluations detailing their performance relative to peers, this has to be taken with a grain of salt. It is very likely that Grok 3 doesn’t outperform its peers in every category, but there is a slim chance these other comparison evals just weren’t run in the optimization for expedience.To start, we can compare Grok 3 benchmarks versus available instruct models.And versus available reasoning models (note how OpenAI’s announced o3 scores exceed these clearly).An important detail, as we’ve seen with OpenAI’s reasoning model releases is, what do the shaded regions on the above plots show? Without exact details, we don’t know the inference cost for each of the models on these reasoning plots. Pushing the frontier in absolute terms is important, but the field overall is getting messier before it’ll get clearer.Regardless, in the above two plots Grok 3 is pushing progress both on standard model training and the new reasoning training. While reasoning training and RL are the hot new things in the AI field, simple scaling and optimization of existing techniques still deliver value.And Grok’s score on ChatBotArena.A model launching at top of every category on ChatBotArena feels like something that should be rare (given it now encompasses many categories like Math, Coding, Style Control, Longer Queries, etc.), but it happened just a few weeks ago with Gemini 2.0 Pro!ChatBotArena is known to favor models that are likely to not refuse requests (we don’t know by how much), as evidenced by Claude 3.5 Sonnet (New)’s relatively low position on the leaderboard relative to its utility, but overall is a hard evaluation to top. xAI’s stated goals of a “based” model should correlate well here.A question we don't know the answer to: How many points of performance on evals do you gain by not caring about safety at all? Internal to the model, i.e. in behavior latent spaces, safety is pretty orthogonal to common high-utility behaviors like reasoning and code, and bigger models tend to do more things without a cost to other behaviors, but there has to be a safety performance margin. Did Grok 3 succeed because of this? It’s too early to tell.At a technical level, Grok 3 is certainly a very big model. We don’t have specifics, but it’s reasonably safe to take a datapoint for scaling still helps for performance (but maybe not on costs). xAI’s approach and messaging has been to get the biggest cluster online as soon as possible. The Occam’s Razor explanation until we have more details is that scaling helped, but it is possible that most of Grok’s performance comes from techniques other than naive scaling.Grok 3’s size to beat existing models feels like when Nemotron 340B beat Llama 3 70B, making it the leading open-weight model at the time, but uptake was slow because the cost relative to the performance gains wasn’t worth it to adopt. We’ll know more about this when Grok 3 is available in their API and we see the exact costs.When models are approximately equal in performance, price and ease of use are the determining factors of adoption.Overall, Grok 3 is a huge technical achievement but not one that indicates a substantial change in who is at the frontier of effective training. xAI is obviously closing in on OpenAI, Anthropic, and most of all Google, but all available data points put these labs ahead of xAI on the frontier of efficient model training. It is good that they are being pressured to deliver more absolute intelligence and not to just continue optimizing their frontier of performance per dollar.Read some other reviews of Grok 3 here and here. Karpathy’s summary is particularly in line with my thinking (while potentially slightly overselling capabilities).As far as a quick vibe check over ~2 hours this morning, Grok 3 + Thinking feels somewhere around the state of the art territory of OpenAI's strongest models (o1-pro, $200/month), and slightly better than DeepSeek-R1 and Gemini 2.0 Flash Thinking. Which is quite incredible considering that the team started from scratch ~1 year ago, this timescale to state of the art territory is unprecedented. Do also keep in mind the caveats - the models are stochastic and may give slightly different answers each time, and it is very early, so we'll have to wait for a lot more evaluations over a period of the next few days/weeks. The early LM arena results look quite encouraging indeed. For now, big congrats to the xAI team, they clearly have huge velocity and momentum and I am excited to add Grok 3 to my "LLM council" and hear what it thinks going forward.Where progress is headingIf these AI models, and the industry writ large, are accelerating, it is important to wonder where they are accelerating toward. Most of the evals we use now to launch leading models are not that representative, in many cases they’re actually 100% out of distribution to normal life. What is the value in solving a competition math problem like AIME or so-called “Google Proof” questions? Time will tell, but the case for usefulness to average users is definitely stretched.Small ChatBotArena improvements are marginal gains in robustness, where something like a 20-point difference in Elo rankings — the relative difference between Grok 3 and the next top model — translates to the model winning something like 51% of head-to-head match-ups. This robustness adds up over time, but it is far from meaning that this model is more intelligent in an absolute sense.In fact, in the case of some of the latest evaluations from the research community, it seems like evaluations are being designed more around being hard than being useful. It is a natural response to models being super powerful to try and find something to challenge them with, but it makes tracking progress and communication far harder.Companies have many internal evaluations that are not shared. Increasing transparency on these would help contextualize what is and is not meaningful progress. Without these, the only benchmark we have for model changes is them becoming more deeply integrated into products. Product-model synergy can enable extremely useful, new workflows, but it makes tracking the progress of AI a proxy measurement.I do personally believe these somewhat arbitrary capabilities we are marching toward will generalize to extended and amplified value, but it takes some "feeling the AGI" to see that these models that are better on esoteric benchmarks generalize to every day use. So far they have. Bigger and generally “better” models have been more robust and easier to find valuable veins in, but we as an industry should be sharing more so that it is not just AI insiders who understand how to track progress.When 2024 was reflected on with meager advancements, evidence is that there was substantial progress but less of it was delivered to users. We only got o1 late in the year and other models were deemed "too big to ship" or the requisite urgency (DeepSeek) did not exist.2025 will be a year of intelligence being put in the user’s hands. The pace of underlying progress with that will continue to be high. The so-called “walls” facing AI progress haven’t materialized, but making sense of the progress we are getting is much more nuanced. Get full access to Interconnects at www.interconnects.ai/subscribe

  --------  

  11:32
• An unexpected RL Renaissance

  The era we are living through in language modeling research is one characterized by complete faith that reasoning and new reinforcement learning (RL) training methods will work. This is well-founded. A day | cannot | go | by | without | a new | reasoning model, RL training result, or dataset distilled from DeepSeek R1.The difference, compared to the last time RL was at the forefront of the AI world with the fact that reinforcement learning from human feedback (RLHF) was needed to create ChatGPT, is that we have way better infrastructure than our first time through this. People are already successfully using TRL, OpenRLHF, veRL, and of course, Open Instruct (our tools for Tülu 3/OLMo) to train models like this.When models such as Alpaca, Vicuña, Dolly, etc. were coming out they were all built on basic instruction tuning. Even though RLHF was the motivation of these experiments, tooling, and lack of datasets made complete and substantive replications rare. On top of that, every organization was trying to recalibrate its AI strategy for the second time in 6 months. The reaction and excitement of Stable Diffusion was all but overwritten by ChatGPT. This time is different. With reasoning models, everyone already has raised money for their AI companies, open-source tooling for RLHF exists and is stable, and everyone is already feeling the AGI.Aside: For a history of what happened in the Alpaca era of open instruct models, watch my recap lecture here — it’s one of my favorite talks in the last few years.The goal of this talk is to try and make sense of the story that is unfolding today:* Given it is becoming obvious that RL with verifiable rewards works on old models — why did the AI community sleep on the potential of these reasoning models? * How to contextualize the development of RLHF techniques with the new types of RL training?* What is the future of post-training? How far can we scale RL?* How does today’s RL compare to historical successes of Deep RL?And other topics. This is a longer-form recording of a talk I gave this week at a local Seattle research meetup (slides are here). I’ll get back to covering the technical details soon!Some of the key points I arrived on:* RLHF was necessary, but not sufficient for ChatGPT. RL training like for reasoning could become the primary driving force of future LM developments. There’s a path for “post-training” to just be called “training” in the future.* While this will feel like the Alpaca moment from 2 years ago, it will produce much deeper results and impact.* Self-play, inference-time compute, and other popular terms related to this movement are more “side quests” than core to the RL developments. They’re both either inspirations or side-effects of good RL.* There is just so much low-hanging fruit for improving models with RL. It’s wonderfully exciting.For the rest, you’ll have to watch the talk. Soon, I’ll cover more of the low level technical developments we are seeing in this space.00:00 The ingredients of an RL paradigm shift16:04 RL with verifiable rewards27:38 What DeepSeek R1 taught us29:30 RL as the focus of language modeling Get full access to Interconnects at www.interconnects.ai/subscribe

  --------  

  39:49

Más podcasts de Tecnología

Podcasts a la moda de Tecnología

Acerca de Interconnects