What kind of company or organization is Reflection?
I think it is ever more important to realize who is releasing models rather than what the models do and how they compare.
Because models iterate at breakneck speed, looking at today's benchmarks is only useful for someone using the models today. Whereas if one builds a product on top of it, or commits to one for a project or team, the company or organization behind it, is far more important. Will they exist in a few months? Do they need a business-model? Are they subsidizing usage with venture capital and how long can they keep this up?
Always glad to see more open-weight models, but this caption on the 2nd demo image had me do a double-take: "Land or Water Generalization Experiment: We recreated the viral X puzzle by asking Beam to create a fixed 180×90 grid for longitudes -179° to 179° and latitudes -89° to 89°, with 16,200 points. This puzzle is a few days old, so could not appear in the training data, thus testing the model’s generalization. Beam gets 95.5% coverage right, putting us between Opus 5 (92.5%) and Fable 5 (97.8%), which shows how well it generalizes to novel new tasks."
Oof, no, this "puzzle is a few days old" is incorrect even if it's a social media trend just recently. Asking a model to generate a world map in this way is _at least_ from August 2025 as it appeared on LessWrong at that time: https://www.lesswrong.com/posts/xwdRzJxyqFqgXTWbH/how-does-a...
Yeah I remember when the original post about this came out. Def not recent. Though I think their point survives in that they didn't exactly RL on this.
Model weights (what is being tested here) don't inherently "access the web" when inference is running. If the model has access to a web search tool, that's a different story.
Maybe that's a rhetorical question but just in case - the search would always be part of the harness. A model is only handling next-token prediction for a given input. That token may be something like [[web search]] to invoke a tool call but the actual call would be handled by the harness.
Yeah it was rhetorical. Search would be implemented as a tool call. Pure intelligence tests would likely have limited tools. But maybe they would have a python sandbox to solve issues like Rs in strawberry.
Hey since it's an open-weight model, I would love to know: how much model safety alignment have you done? Did you do any sort of post-training and what restrictions are in place
Ideally i would like to place my own restrictions and align from scratch, currently I am resolved to do harness alignment using tools like Prismor but would love to do my own post training alignment
I'm on the waiting list... Couldn't find any download option, so I suppose it is only obtainable through their API. Strange way to distribute open weights model.
I feel like this is a marketing miss. If they had held their announcement until the model was released, I would have grabbed it and started running it through my benchmarks. It probably doesn’t get a place in the rotation based on their own description of its performance, but now the weights live on the server, I’m probably following them on HF and I will remember to check in every time I ls the models folder. With the announcement only, none of that happens and I’m likely to forget about this by the time it actually gets released.
The email harvest move just doesn’t fit where we/they are in the cycle. There are established players and a buffet of models to choose from (plus a ton of empty hype). The first move at this point for any new entrant should be to show, not tell. Even an API only release with the promise to open weight would be better (actually probably all around better since most people can’t run this locally).
I wish this lab and all the labs releasing the best. It’s a brutal landscape to sink millions of dollars into for a guaranteed “behind x model from a year ago” evaluation. But, I do believe there is genuine innovation left to uncover.
> Beam is a sparse Mixture-of-Experts model with 501 billion total parameters, 23 billion active, built for coding, reasoning, and agentic workloads.
> Beam’s capabilities come from major investments in both pretraining and reinforcement learning (RL). We pretrained the model on 23.8 trillion diverse, curated, high-quality tokens from the web and proprietary licensed datasets, matching or outperforming available similar-sized open base models. In parallel, we developed the algorithms, training environments, and infrastructure needed to sustain high-compute RL at exceptional scale. Our high-compute RL run generated over 100 million rollouts on 10.5K NVIDIA GB300 GPUs over 4 weeks of training.
Early access, no weights no tech details, just a sign up here for info
I'm all for more open models, but talk is cheap and this is a rather pointless announcement without anything backing it up. Publish your weights and HF repo or shut up IMO.
Is that all that a company about to give away the product of 10,000 GPUs running for a month gets to be now? give it away without a single promotional post or shut up? I support open source as much as the person but this is pretty caustic.
I am also starting to take issue with "we've developed a new cutting edge model, and nobody can use it" announcements. Most recently with Google's Argon, at least they were using it internally and they'd be slowly rolling it out. This one is from a company I've never heard of, they're not releasing weights or offering API access at this point, it feels like a fairly worthless announcement.
> give it away without a single promotional post or shut up?
I think the point is that people are happy to see promotional posts when they actually release it, but only then and not before.
Unfortunately pinky promises from corporations to release something at some indeterminate time in the future aren't worth the bytes they're stored in, especially in the AI industry which is full of grifters and charlatans.
Yes -- because they're late to the party (high performing open weights models have been a thing for a couple of years now), and therefore will be compared with all the other open-weights model providers they are competing against.
It's not just that they are doing users a favor with weights; they are just as much seeking favors with attention and usage (in a crowded market!).
The open weight community really is an odd one. Millions of Dollars for pre and post training given away for free and most often with very permissive licenses that allow commercial use (be it US, EU or mostly Chinese)
Yet going by the comments on localLlaMA or HN, those companies are the devil :-D. Colour me surprised.
You know, the other night I trained a model on my secret stash of GPUs, that now outperforms Opus 5.5 on pelican benchmark and Jev on classification speed, while running on a potato.
And also a "proprietary data set" hahaha... Probably just means they don't want to show it, and it is data, that either they shouldn't have, or that there is nothing special about their training data and it is just meant to sound like there is some secret ingredient, while there is none.
Not sharing the data is pretty standard because 1) it tends to get the lawyers involved and 2) good data is critical for getting good results.
Imo you can get better results with great data and generic modeling techniques than with incredible modeling techniques and crappy data. Because if you have crappy data, you won’t even know if your model is good because your evals will also be bad.
This is why Anthropic is throwing a fit about the Chinese distillation “attacks”. Clean reasoning traces are gold.
Companies pay lots of money for proprietary agentic trajectories which are used during RL. These are things like "Task: summarize stock levels for months end accounting" which then traces the task though using SAP to look at different SKU stock levels, exporting them and generating summary Excel spreadsheets.
This is very different to the "scrape the internet" datasets that a table stakes for training a LLM.
Data has copyright issues, so one can't share it generally without getting permissions from all of the copyright holders. The data is not theirs to share, anyways. The derived (learned) weights are a different matter.
True for the pre-training data scraped from diverse sources. Less so for the later stage data for RL which is by all account more of a differentiator. In most cases the labs themselves produced the data so they are the copyright holders (or they are borrowing it from other labs via distillation). A lot of it is synthetic data, and since you can't copyright AI output, it becomes less about copyright and more about trade secrets.
this is very normal for frontier lab companies. you need good data either synthetic or labelled (all the chinese open source models have their own armies of data labelers)
Bigger and still worse than existing free Chinese models that are smaller? Open weight models are nice, but at this point it seems western models are very far behind Chinese ones, despite Chinese companies publishing a lot of their findings. I hope we get more open models and more providers, as being stuck with a model from China or US with no competition is risky.
Google does do a great job with Gemma models. It's one of the few language models actually good at language. OpenAI's top closed models can't even write norwegian correctly.
It takes time / few iterations to get it right (and it's moving target), but yes, expensive trial, my personal feeling is that they went a bit too high, at the same time who knows, maybe good move – as they're saying RL didn't plateau. It feels like they had something like $100M budget for it?
The world will be a better place when we stop espousing Chinese-anything. They don’t treat their people well, they don’t care about them, just cogs in a machine. I’m completely disgusted with how hn people offers China up on a platter like they’re an example of something to emulate, I wish I knew how we got to this point.
I hear this so much from westerners who have no connection to the people or the country.
-China passed a law forbidding companies from replacing workers with AI.
-China regularly punishes CEOs and corrupt government officials who do bad things.
China promotes open source to the world enabling everyone in every country to have equitable access.
From living in china for over 20 years and talking to many many people, the majority is happy with the trajectory of their country.
Could any for these be said of the west?
I won't get into specifics on which countries regularly throw bombs on children, use starvation and blockades as weapons, and completely disregard the massive dissatisfaction of their citizens...but it sure isn't china.
Please clean your own house first before you complain about your neighbor.
This sentiment was really common circa 2005, or at least general anti-China rhetoric, where I lived in the Tri-state. Funnily enough, I have never felt more a cog, living here in the U.S., than I do right now. Maybe you are right, but my guess is wherever you live, yours is a bit of a glass house as well.
Yeah, we should only promote, like... Finnish, Norwegian, maybe Swiss products, like Apertus. Not this unethically-made baby oil[1] from repressive torment nexuses of USA or China.
It's pretty clear from their framing ("Beam advances the Western open-weight frontier") that one of their main selling points is not being a Chinese lab.
I can't imagine that mattering to many individuals, but I guess someone out there has a government contract that forbids the use of foreign models
Sometimes we just enjiy having a conversation with people. It's how we got many answers before Google exists. I believe such choices have many, positive effects on people that society is losing.
because my templeos goes straight from ring0 after bios straight into a ui for hackernews that only lets me scroll, click into comments and type comments.
Multiple independent approaches are cool and all but fully open source model training (datasets, pipeline, checkpoints) should be taking advantage of being open and share runs/budget between different entities.
I disagree. Sure let them play and see if they can improve. But this model has more compute and more training data than the predecessors it fails to surpass. That only means their training regime is inferior if their predecessors did so much more with so much less. That inferiority should not be encouraged.
You don't just magically do better than everyone else on every metric on your first go at something. Doing worse than others and refining is how pretty much everything works.
The reality is they trained a model and it looks worse on benchmarks than Qwen or GLM. I don’t see how sharing the weights hurts anyone? Even when Llama 4 came out and it was a dumpster fire, it didn’t affect me personally.
> That only means their training regime is inferior if their predecessors did so much more with so much less
Hard to imagine how that wouldn’t be the case. They probably missed the boat on distilling Claude (or their lawyers said no), they probably didn’t hire an army of math PhDs to write reasoning traces, they don’t have millions of DAUs in a coding agent to train from, and they probably have less money, less experience, fewer top tier researchers, and fewer resources for experiments. They are an underdog without a doubt.
None of that means they shouldn’t release their model.
Them releasing the weights doesn't hurt anyone. It's the peanut gallery clamoring to put them onto the same pedestal as actual tier 1 companies simply because they aren't named openai or anthropic that is hurtful.
Openai, grok, and Anthropic aren't distilling. Theyre just second class. It's not a big deal, we just shouldn't be lauding them for being second class.
And it still sucks. My apologies to the Cursor team but that's just very very poor performance.
Alternative explanation is that the Chinese have far more technical talent than anyone else, along with the infra and capital to build out these models.
Reflection is explicitly marketed as the 'US' DeepSeek
seems like they are aiming to provide both inference and RLaaS for american companies and western govts. even if they never fully beat deepseek if they get close enough the fact that they're American will help them close deals
I remember being in the room with pretraining day 1 to help monitor the training job launch. Watching this model train from day 1 has been an amazing experience!
Outside of ML metrics, you're monitoring the health of every piece of hardware in the system. You need to make sure that you have every GPU, every CPU, the PCIe buses, the networking fabric are all working without any errors. You need to ensure that you can respond as fast as possible to any possible error. One bad component can bottleneck the entire job.
I really enjoyed reading the log book from the training of OPT-175B at Meta… I guess it’s all classified info but it’d be fun to read a blog post about the crazy day to day issues you run into when doing things at this scale :)
GPU failures are frequent enough that at a certain scale, you constantly have workers dropping out. Designing systems that can still keep training is very interesting!
Does this one also routes to Claude under the hood like Reflection 70B did? I recall they even run a basic regex to remove "Claude" from the output. Then they promised to be completely transparent on the postmortem (they claim they had no idea what had happened), but the postmortem never came
Releasing open-weight models at this scale is a massive milestone for the community. Access to transparent model internals is foundational for trustworthy systems.
I don't find open-weight models that impressive anymore. MiMo-V2.6 already showed that you can have a not so crazy architecture and enough compute, the bottleneck is then just the data. OAI and Anthropic are largely the frontier models because of the synthetic data they made. They have a large customer base and have the user's data as well as knowing what tasks their customers use the models for and what domain they should get synthetic data for.
Interesting. Have heard about the need to create synthetic data for LLM's, but didn't realize it was such an important factor. Although, how big a factor synthetic data is the next question, but guess there is no way to truly verify how much difference it makes with these closed models.
This is literally what the labs have been doing over the past year. The whole idea of emergence is a lie, there is some interpolation and superhuman long evaluations the models can do, but almost all the gains are from synthetic data. They hire thousands of professionals and pay them as much as 200$/hour to create many tasks that they want the model to perform and use these to teach the model on how to do it with RL. OAI had 30,000 contractors from Mercor for Sol 5.6.
When you see Opus suddenly becoming great at blender or some other 3D graphics, that's because they hired professionals and had them do similar tasks that people are looking for. They keep having better professionals at each iteration and so the quality improves. There is no emergence or "General" intelligence. The model doesn't learn to become better at a task because of scaling laws or emergence or whatever they might wanna say, it is literally RL on tasks that they want the model to perform well on.
Also interesting. Do you happen to know if once they've used those thousands of contractors to teach the model something like Blender (or some similar app) using RL, when they training a new model, do they need to use same contractors again to teach that new model the same behaviors?
The reason I ask is even though it can take hundreds or thousands of contractors to teach a model a certain behavior, wonder if they really only need to do it once for each desired behavior? (of course, future models might expand and refine this previous training) Because if that is the case, then wow, then future models can really expand their capabilities really very fast.
...Am wondering if they somehow record a training session so they can play it back whenever they need to train a new model with the same info? Or maybe the new models can just use distillation from the old model to relearn the old behaviors?
The contractors don't directly teach the model. They create datasets. Mostly they create tasks within an environment that the current generation of models wouldn't be able to do, they then write maybe a solution, a set of rules for evaluating the response, and whatever is needed for the RL. These tasks form a dataset. You can see examples of a task in the Mimo dataset that was open-sourced recently. The dataset would then be used by engineers for post-training of whatever model. Some of the model iterations that are released every month tend to be just a further post-training of the same base model that was pre-trained months ago. OAI recently has been doing a lot more pre-training, but for a long time they had the same pre-trained base model. This is why you see so many releases done so fast by the labs, they just need to post-train the same base model on whatever task they think would be better suited, I would speculate based on what users want and what the benchmarks test for.
Regarding your second question, I think if they want to further post-train a model using new data, they wouldn't feel the need to re-train it again on the data that it has already being trained on. But you never know. If the model is a completely new pre-trained base model, then they could either train the model using all the data and/or use a previous model to teach it. There is definitely a bunch of tricks they do to evaluate the models and check the performance or whatever their recipe is. It's really up to what the engineers would prefer. But you get the core idea, the models are not suddenly coming up with how to use the Blender on their own, they are explicitly being trained on a dataset curated by a professional that teaches the model how to use Blender. Surely there is another aspect that if the model gets better at coding, then it also helps it become better at Blender, and you have that transfer learning. However, there is no emergence or a deity popping up. But you get people who were evaluating theses models on blender use and suddenly seeing the model ace their tasks and they think they are dealing with a super-intelligence. They then undergo an AI psychosis once they try and extrapolate the (super)-exponential improvement in that one task over the next few months and across all other domains.
Regarding your third question, I already answered at it. But, when it comes to training, they definitely freeze the weights after each run just in case an issue arises and they need to address it (a GPU not working or the loss value blowing up).
Is it worse than the top open-weight Chinese models? Yes, it is, but at least the West has joined the party, and hopefully they will iterate on this and keep up the pace. The Chinese labs will certainly release new and powerful versions soon, so it's all about relative pace right now.
Beggar choosing: my kingdom for more 90B-133B MoE local models. Especially with disk offload, that is a function/performance sweet spot for Mac workstations with 64GB-128GB of RAM.
I am really curious why after Qwen-3.8-flash and deepseek-v4.1-flash newer open weight models (or proprietary but they don't tell) don't use n-grams. It looks (to me) like they are a cheep way to add more knowledge to the model.
What do I mean by cheap? You can rely on the SSD to retrieve the relevant tokens as no computation is needed meaning you can leverage storage (or cpu ram if you don't have unified memory) to serve part of the model which (to my understanding) is much cheaper to get than GPU RAM.
Anyone know what am I missing? Or is it that the pace of iteration for labs slow enough that they can't actually leverage it yet?
Instead of yet another mediocre but fully-made-in-the-West open model (alongside Mistral, Trinity, Poolside, Inkling, etc etc) I'd really love for a Western neloab start the same way Qwen did: by focusing on post-training. Qwen's first release was a Llama 1 finetune [1]! Once they made it useful, they started working their way back in the stack to also do their own pretraining, etc. Starting with pretraining feels like such a waste: there's millions of dollars of crystallized compute and data sitting around in the Chinese model weights. Why not start with one of those, and only work your way back to pretraining once you've released something you can prove is useful?
At least for RL, you don't need a base model — the rollouts are run in an inference engine with an instruction-tuned model using a chat template. You can start with just that!
Yeah but they are already RL'd too. No doubt you can further improve an RL'd model by doing your own better RL on top of that. But it's not going to be the same as starting with a base model or instruction tuned model.
The meaningful test begins after the weights arrive: fixed harness, network disabled, repeated trials, and real latency, memory, energy and cost per completed task.
It's true that no benchmark communicates the whole picture, and we won't really know how it behaves until weights are out, but the performance here doesn't seem particularly groundbreaking just based on the benchmark.
Very little in terms of the layers they use. Calling it now, they are using global layers everywhere, making the model basically unusable due to high KV Cache use.
the performance chart puts the better open source models behind the fold making it seem like it outperforms them... but it doesn't! all for open source models but this announcement is misleading
Any time a new lab shows up, folks complain about how their models are worse. Really? It would be nice if a new comer comes from no where and beats everyone, but that's rarely the case. The good thing is that other labs/people are figuring out how to build this, and if they keep at it then this is as bad as it gets for them and it would hopefully get better. A new entrant to the market is good for everyone.
honest question: how honest do you think people are about their improvements and performance compared to objective results when all you do is praise them?
Nitpicking but I really wish this benchmarks table were easier to read. Should show which columns win in each row and should not require horizontal scrolling to see across.
ok so they are comparing themselves to and claim to be beating GLM's last generation GLM 5.2 model, GLM 5.3 Flash is a monster, this is honestly embarassing
Open model that is not yet open or widely accessible via API. Primarily comparing to non-SOTA models like Inkling and GLM 5.2. Included comparison to GLM 5.3 and DeepSeek V4.1 Flash in the table, but not in the charts (I assume they would make them look bad). Also no results from AA Index or Arena.
I am a great fan of open weights model, but off late I am starting to loose track on the capabilities of the latest models released and now a days most of the open weights models are above 100 B params which is not going to run in our laptops. What happened to Jev hype? A 501B Beam model is not going to help with it (Non AR Schema driven responding under 1 second).
I am more interested with a SOTA Frontier 8B-10B model. Is this even possible?
very curious to see more about what kinds of hardware you can run this on and the perf. characteristics… on the face of it, it seems like optimizing for inference speed might(?) be good for running on smaller hardware, but i suppose it could be the other way around and it is actually much resource-hungrier for the number of parameters, etc. …
If you don't buy into "America good, China bad" narrative, this new entrant & release by Inclusion Ai is a lot more exciting by every measurable metric.
All of those are good references. Other folks in the thread are missing distinctions between pre training (~the internet + curated sources) and post training (~instructions and RL)
If you ask a model, they will generally tell you where to get data. Modern frontier models have the large advantage of having tens if not hundreds of millions of users providing use cases to train against to improve their responses.
Note that this sort of distillation is NOT for pre-training data (which is tens of trillions of tokens). I think the allegations against Chinese companies by Anthropic is more so that they distill SFT data (which is good for post-training, but you still need a strong base model)
pretty wild that, per public sources, Reflection AI raised over $4 billion and hit a $25 billion valuation while operating in total stealth, without ever releasing a single public product until now. Beam (501B) seems be their first-ever model drop. Or am I missing something?
I think it is ever more important to realize who is releasing models rather than what the models do and how they compare.
Because models iterate at breakneck speed, looking at today's benchmarks is only useful for someone using the models today. Whereas if one builds a product on top of it, or commits to one for a project or team, the company or organization behind it, is far more important. Will they exist in a few months? Do they need a business-model? Are they subsidizing usage with venture capital and how long can they keep this up?
Is reflection a company? University lab? NGO?
Oof, no, this "puzzle is a few days old" is incorrect even if it's a social media trend just recently. Asking a model to generate a world map in this way is _at least_ from August 2025 as it appeared on LessWrong at that time: https://www.lesswrong.com/posts/xwdRzJxyqFqgXTWbH/how-does-a...
one would hope that they disable websearch and internet access (maybe all tools?) when doing generalization testing?
Ideally i would like to place my own restrictions and align from scratch, currently I am resolved to do harness alignment using tools like Prismor but would love to do my own post training alignment
The email harvest move just doesn’t fit where we/they are in the cycle. There are established players and a buffet of models to choose from (plus a ton of empty hype). The first move at this point for any new entrant should be to show, not tell. Even an API only release with the promise to open weight would be better (actually probably all around better since most people can’t run this locally).
I wish this lab and all the labs releasing the best. It’s a brutal landscape to sink millions of dollars into for a guaranteed “behind x model from a year ago” evaluation. But, I do believe there is genuine innovation left to uncover.
> Beam’s capabilities come from major investments in both pretraining and reinforcement learning (RL). We pretrained the model on 23.8 trillion diverse, curated, high-quality tokens from the web and proprietary licensed datasets, matching or outperforming available similar-sized open base models. In parallel, we developed the algorithms, training environments, and infrastructure needed to sustain high-compute RL at exceptional scale. Our high-compute RL run generated over 100 million rollouts on 10.5K NVIDIA GB300 GPUs over 4 weeks of training.
Early access, no weights no tech details, just a sign up here for info
I think the point is that people are happy to see promotional posts when they actually release it, but only then and not before.
Unfortunately pinky promises from corporations to release something at some indeterminate time in the future aren't worth the bytes they're stored in, especially in the AI industry which is full of grifters and charlatans.
It's not just that they are doing users a favor with weights; they are just as much seeking favors with attention and usage (in a crowded market!).
I think the comment you are replying to is unnecessarily hostile too though.
Yet going by the comments on localLlaMA or HN, those companies are the devil :-D. Colour me surprised.
Will release weights soon.
That means that they have the 31th of October as the deadline to make true their claims.
The fact that they give early access to some may mean that they want some beta testers before the public release.
Imo you can get better results with great data and generic modeling techniques than with incredible modeling techniques and crappy data. Because if you have crappy data, you won’t even know if your model is good because your evals will also be bad.
This is why Anthropic is throwing a fit about the Chinese distillation “attacks”. Clean reasoning traces are gold.
Companies pay lots of money for proprietary agentic trajectories which are used during RL. These are things like "Task: summarize stock levels for months end accounting" which then traces the task though using SAP to look at different SKU stock levels, exporting them and generating summary Excel spreadsheets.
This is very different to the "scrape the internet" datasets that a table stakes for training a LLM.
Xiaomi released a fairly developer-centric dataset like this here: https://huggingface.co/datasets/XiaomiMiMo/MiMo-V2.6-RL-oss
SpreadsheetRL is another fairly specialized dataset: https://spreadsheet-rl.github.io/
> We will release the weights, technical report, model card, and developer artifacts later this month.
Google does do a great job with Gemma models. It's one of the few language models actually good at language. OpenAI's top closed models can't even write norwegian correctly.
[1] https://i.imgur.com/rJSG019.jpeg
Am I missing something?
It's pretty clear from their framing ("Beam advances the Western open-weight frontier") that one of their main selling points is not being a Chinese lab.
I can't imagine that mattering to many individuals, but I guess someone out there has a government contract that forbids the use of foreign models
InB4: kids these days :shakes-fist-at-cloud:
Asking an easily-searchable question is just lazy.
/s
The conversation is why.
500B params performing worse than other OSS of the same size is pretty meaningless if no one will use it.
> That only means their training regime is inferior if their predecessors did so much more with so much less
Hard to imagine how that wouldn’t be the case. They probably missed the boat on distilling Claude (or their lawyers said no), they probably didn’t hire an army of math PhDs to write reasoning traces, they don’t have millions of DAUs in a coding agent to train from, and they probably have less money, less experience, fewer top tier researchers, and fewer resources for experiments. They are an underdog without a doubt.
None of that means they shouldn’t release their model.
Says who? We know Grok does at the least. They admitted it openly.
Alternative explanation is that the Chinese have far more technical talent than anyone else, along with the infra and capital to build out these models.
My money is on the latter explanation, tbh.
seems like they are aiming to provide both inference and RLaaS for american companies and western govts. even if they never fully beat deepseek if they get close enough the fact that they're American will help them close deals
I really enjoyed reading the log book from the training of OPT-175B at Meta… I guess it’s all classified info but it’d be fun to read a blog post about the crazy day to day issues you run into when doing things at this scale :)
You're talking about real data created and curated by humans to help in training LLMs.
The reason I ask is even though it can take hundreds or thousands of contractors to teach a model a certain behavior, wonder if they really only need to do it once for each desired behavior? (of course, future models might expand and refine this previous training) Because if that is the case, then wow, then future models can really expand their capabilities really very fast.
...Am wondering if they somehow record a training session so they can play it back whenever they need to train a new model with the same info? Or maybe the new models can just use distillation from the old model to relearn the old behaviors?
Regarding your second question, I think if they want to further post-train a model using new data, they wouldn't feel the need to re-train it again on the data that it has already being trained on. But you never know. If the model is a completely new pre-trained base model, then they could either train the model using all the data and/or use a previous model to teach it. There is definitely a bunch of tricks they do to evaluate the models and check the performance or whatever their recipe is. It's really up to what the engineers would prefer. But you get the core idea, the models are not suddenly coming up with how to use the Blender on their own, they are explicitly being trained on a dataset curated by a professional that teaches the model how to use Blender. Surely there is another aspect that if the model gets better at coding, then it also helps it become better at Blender, and you have that transfer learning. However, there is no emergence or a deity popping up. But you get people who were evaluating theses models on blender use and suddenly seeing the model ace their tasks and they think they are dealing with a super-intelligence. They then undergo an AI psychosis once they try and extrapolate the (super)-exponential improvement in that one task over the next few months and across all other domains.
Regarding your third question, I already answered at it. But, when it comes to training, they definitely freeze the weights after each run just in case an issue arises and they need to address it (a GPU not working or the loss value blowing up).
What do I mean by cheap? You can rely on the SSD to retrieve the relevant tokens as no computation is needed meaning you can leverage storage (or cpu ram if you don't have unified memory) to serve part of the model which (to my understanding) is much cheaper to get than GPU RAM.
Anyone know what am I missing? Or is it that the pace of iteration for labs slow enough that they can't actually leverage it yet?
1: https://en.wikipedia.org/wiki/Qwen
It's great to see a company that acknowledges it still needs improvement instead of making false claims.
The price of that kit is rapidly approaching $10k.
It's true that no benchmark communicates the whole picture, and we won't really know how it behaves until weights are out, but the performance here doesn't seem particularly groundbreaking just based on the benchmark.
participation awards are not helpful.
Access is currently limited. We'll contact you if early access becomes available.
It's interesting how the industry converged to this very term, given that very less work is being done by horses since quite a while.
I am more interested with a SOTA Frontier 8B-10B model. Is this even possible?
https://github.com/inclusionAI/Ling
BTW, this is also an AI lab from China
Where do I get the data?
I mean, this many models. They have to start somewhere.
e.g. fineweb dataset is 50TB https://huggingface.co/datasets/HuggingFaceFW/fineweb
I recommend checking papers from Datalogy, Nvidia Nemotron, Ai2 (Ollmo, Tulu, ...) and the recent model from Aleph Alpha if you want to learn more.