Is there a reason these pelicans always have roughly the same composition (side-view, 2d, biking right, flat ground beneath, etc)? I don't see any of that detailed in the prompt, yet they all seem to generate roughly the same image of differing quality.
I was going to ask the exact same question earlier but deleted it after thinking “I’m sure Simon has done some sort of discussion on this.” Since it does seem novel to you, too, it would be really interesting to read more about this phenomenon.
It's my impression that it's common in western culture, where text is read left to right, and timelines are visualized as going from left to right, to also animate things going from left to right, since westerners thus have an instinct that "right = forward", so it "feels right" (familiar). I wonder to which degree this is reflected in the training data? And if you'd be more likely to get left-facing pelicans if you prompted it in Hebrew, Arabic or another right-to-left language?
Someone studied this (among other thigns): https://dylancastillo.co/posts/pelicanmaxxing.html . Pelicans on bikes always face right in this test, but other animals on other transportation methods sometimes face left.
The more generic your prompt, the more generic the response. It's a regression to the "mean" of the training data aka GIGO for AI.
It's like when you ask your average person off the street to draw a house - it'll almost always be square with a triangle roof, one door, and two windows.
In the pelican/bike example, it's probably a bit of a self-perpetuating snowball too. If the earliest examples were bike left-to-right, flat ground, etc. then they are also being scraped up in future LLMs.
I sincerely believe I've never had a single original thought™ in my whole life.
There is this scene in the HBO series Westworld where a "host" says some words in sequence which is shown on a display as she says it. Of course, even me thinking of this scene and connecting it to your comment was not original, someone else clearly had the same programming as me.
A medium blog post says
> Pair what with me?” — the moment Maeve (a humanoid android) uttered those words in Westworld (Season 1, Episode 6: “The Adversary”), something clicked. Not for the average viewer, but for me, a STEM educator and AI enthusiast who, just weeks earlier, had read Stephen Wolfram’s seminal essay, What Is ChatGPT Doing … and Why Does It Work?
It's not even that old - but back when it was aired, an AI that can not just string together coherent sentences, but produce coherent reactions in novel, fully unintended contexts, like Maeve was doing there? It was totally a sci-fi premise.
Now we have AIs capable of that and more, and no one bats an eye.
Indeed: “Our hosts began to pass the Turing test within the first year.”
Required sci-fi suspension-of-disbelief in 2017, and then at some point in the last few years we just blew by that one.
Later seasons of the show were much less dramatically satisfying, but also played out the consequences of the science of artificial intelligence demonstrating as a side-effect that human intelligence and free will might have as much of an uncertain foundation as that of machines.
How much data from the Panopticon, how many parameters would it take to train a model that could predict your responses?
Tesla had the same thought. He called himself an automata: "entirely controlled by the forces of the medium" It inspired him to create the first remote control vehicle.
It's just the simplest most recognizable form of a house. Like how a smiley face is so generic and simplistic but everyone will know what it represents. Just two dots and a line yet it's easily and unambiguously understood to represent a human face and a happy emotion.
Search Google Images for "bicycle". Almost all bicycle product shots are staged the same way: side view, going left-to-right. It makes sense to me that given that skew in the training data, the model grounds itself in the bicycle.
and furthermore, this is because the drivetrain is ~always on the right side of the bike - if you want to inspect or admire a bicycle you look at the right side, as you might look under the hood of a car.
(Why the drivetrain is on the right, I don't know. But most bike parts follow open standards so it's quite entrenched.)
> and furthermore, this is because the drivetrain is ~always on the right side of the bike
While I'm sure this factors into things for advertisements for bike components, there is also just a general preference that westerners have for left-to-right motion. Not just in bike ads, but all ads with (or suggesting) movement. And also not just ads, but movies where directors believe left-to-right motion is associated with progression and right-to-left motion is regressive.
Since most languages read from left to right, rightward movement tends to read as forward progression. So when showing a bicycle in side profile, having it face right feels more naturally like it’s moving forward.
I can’t tell you why it’s always on the right, but it’s always on the same side because of network effects.
Bicycle frames are not fully symmetric left-right because you need things like a mount point for the derailleur hanger, and optionally affordances to keep the chain off the stays when the wheel is removed.
Those things have to be on the same side as the chain. Bikes designed for disc brakes additionally need a mount point for the brake caliper on the opposite side from the chain.
Additionally, rear wheels are not symmetric: the spokes on the chain side connect to the hub closer to the plane of the rim. That is, they are more perpendicular to the wheel’s rotational axis than spokes on the opposite side (which is why you should always mount a single pannier on the chain side). This asymmetry is to provide space for the gears.
So once the industry decided to put the chain on the ride, you can’t very well make a group set designed for a left chain if you want it to work on the vast majority of frames.
Yes, I do a thing where I ask the machine to generate responses in the form of a lizard talking to a cat. The lizard is always a green gecko and the cat is always orange, which I never specify.
Wow, I actually had this exact idea. I was specifically curious as to how well a given LLM could understand a DSL that hasn't changed much in a couple decades and doesn't have nearly as many examples to learn from online. Seems like it did alright, all things considered.
Ohh, horizontal wheels. They’re about as good as I expected, models have pretty bad spatial awareness. I would expect Fable to be a bit better than old models, though.
I wonder how a multi-modal model would do with a harness and tool calling? Specifically a "render" command that produced an image output enabling it to iterate. (Well I see you did this manually with gemini 2.5 pro but I still think it would be interesting to explore various harness setups.)
> GPT-5.1 Codex
> monstrosity
What are you talking about? That's clearly a sci-fi pelican on a hoverboard (successor of the humble bicycle) wearing a visor. Truly visionary.
The canonical view of a bicycle is facing right. Usually, people want to draw/photograph/depict the side of the bicycle with the running gear, which is on the right side of the frame for historical reasons.
The thing that distinguishes pelicans from other birds does so most strongly in profile. If you're looking straight at one, the throat pouch would be hidden by the beak.
I bet if it instead had something to do with black widow spiders we'd find that we're most often looking at the bottom of the spider's abdomen, regardless of whatever non-spider-like activity is supplied.
well it is svg, it is doing it from circles and lines as primitives, it wants to do it simply and kind of builds the whole thing hierarchically. Making it 3d is way more complicated (as the POV example shows) and the prompt doesn't say 3d anyway
Yes. It's because you are asking it to generate an image of a pelican riding a bicycle. If someone asked you to draw a pelican riding a bycycle, would you interpret that to mean using 3d photorealism? LLMs follow conventions. The convention for an animal riding a bike is to create a childish 2d line drawing.
I wonder, given Simons reputation in AI benchmarking, whether model providers try to train or tweak their models to perform better at drawing bicycles and pelicans?
It would be very embarrassing for any lab to benchmaxx the pelican on bicycle svg prompt, since it would be very easy to detect it by varying the prompt.
The amount of discussion around it means that the test and all the reviews of results, images, approaches etc are implicitly included in training data.
It’s not deliberate “benchmaxxing” but things that are discussed a lot online are naturally things that LLMs learn better.
Did any LLM so far draw pelican knees correctly and have them bend in opposite direction from human knees? Knees of many animals bend opposite to humans.
Did any LLM draw the front bicycle wheel correctly? ie. center of front wheel slightly AHEAD of steering wheel axis. This is done for bicycle stability.
I was interviewed for a job as a software developer last week and they asked me to draw a picture of a pelican riding a bicycle.
Aced it, got the job as a senior software engineer.
The interviewers afterwards said "it is SO refreshing to find a software developer who actually knows how to code - never seen such a high performance focused, well built pelican on a bike - you have the skills we need".
I interviewed as a software developer at LinkedIn. The interviewer asked me to demonstrate my prompting skills, so I had AI write an article about what the recent death of my father taught me about B2B SaaS. Reading it brought tears to his eyes so he hired me on the spot.
"I was interviewed for a job as a software developer last week and they asked me to draw a picture of a pelican riding a bicycle. Aced it, got the job as a senior software engineer."
That is the best joke I have heard this year. Ready for a stand-up comedy special. Or a song. Superb!
At this point the only thing they're useful for is visualizing the differences between effort levels and roughly tracking the progression of models within a specific model family. And they still do that really well!
is there a reason there are so many common base decorative elements across pelicans on bicycles? For instance, there's a lot hats/helmets and scarfs/capes across models.
Used Muse Spark 1.2 and was not impressed at all. Fast and cheap but even GPT 5.6 Terra felt much more capable. Also not really looking to support a company that was just forced to pay $18B for mental health damages.
I'm party using 1.2 to reverse engineer and re-implement an old game binary and it has been quite good and fast. The contributor pricing is very attractive, excited to try 1.3 and see if I feel a difference. 1.2 can get stuck outputting similar sounding thought summaries with no apparent progress when asked to solve bugs. Then I've switched to GLM-5.3-Flash which for this use case has been clearly better at finding suspected causes and following tracks.
Meta has an enormous amount of compute. They are either going use it making and inferencing models or they are going to sell their excess capacity to model providers. Zuck had to completely rebuild his AI team after the Llama 4 launch mess.
Progress is iterative. Everyone is always riffing on other’s ideas and can execute on them given enough support (eg $$). The person to get to an idea first is just 5% away, so it’s possible to catch up.
Moreover,I think it’s impossible to know if you’re hitting a portion of the sigmoid, because there will often be an idea that changes the trajectory altogether.
In 2024, there was a ton of talk about the plateau. Reasoning was an iteration on chain of thought, but it didn’t really work. Deepseek proposes RLVR as a way to get around the lack of $ they have to produce human reasoning trace data. That small iteration catches the eye of OpenAI and Anthropic, turns out to be way more important than even DeepSeek could have ever expected when it comes to improving LLMs for coding, and last 18 months have been an exercise on riding that insight to the nth degree.
That one small iteration brought us a lot of progress. Now we’re seemingly exhausting the impact of that one insight, but there may be another soon enough.
openai did human crafted chain of thought dataset training. deepseek didn't have the resources so they attempted RL. doing RL correctly is hard because of the risk of model collapsing.
Totally, RLVR as a concept predates DeepSeek; but they proposed a version that was simple and scalable. Popularizing a specific version of a technique is exactly what I mean by iterations on a theme. It’s only 5% different from what others tried before, but that 5% difference showed a lot more potential than other versions of the same idea.
Since DeepSeeks GRPO, they’ve been improvements as well like AliBabas GSPO that have gotten wide adoption. Again iterations
Even if all the big ideas are gone and we are entering a new part of the curve, there is still an enormous amount of improvement possible. Just iterating on data mix/quality etc, training pipelines, reward functions, specific ways of reasoning (which i guess is mostly just data still) for the next 20 years will yield a looooot. And that's just the models. The harnesses/application layers/whateveritgetscallednext space has 20 years of progress to make.
So one model is "Not used to improve our products" and is 10-20 times more expensive to the "Used to improve our products"-model.
Given this is Meta, my immediate assumptions that one is cheap because it lets me "be the product". I know I'm rushing to conclusions but there is zero trust here. The brain will do its thing. And the wording here is giving the brains a lot of wiggle room.
Given OpenAI and Anthropic's behavior, do you really expect them to be singled out for this practice? Zero trust has been in "LGTM" territory for years now. Meta's bet against people taking a principled stance arguably paid off great.
What is the confusion? They directly state that you are the product if you use their discounted offering. It isn't an assumption that should lead you to this, it is Meta's very direct communication that should lead you to this
I think it's more that the "not used to improve our models" is expensive because companies need that. It's simple price differentiation.
In other words, it's not that Meta really wants your data and they're willing to pay top dollar for it. It's that companies really don't want Meta to have their data and they're willing to pay top dollar for that.
The previous version was, in my experience, the best free model available on OpenCode. It's been very good at simple/moderate tasks where I am precise in my ask and it doesn't need to make a ton of undefined assumptions. Hopefully this new version is also available on opencode for free.
By default, even without the training endpoint the pricing is pretty competitive, especially against Opus and Fable. [1] The 'muse-spark-1.3-contributor' endpoint is by far the cheapest, significantly cheaper per M than ChatGPT Luna, significantly smarter than Luna too.
This price/intelligence beats even legacy DeepSeek V4 Flash pricing.
I'm wondering whether anyone has yet extracted AWS keys from a model trained on user input. Because users are definitely feeding secrets into these "contributor" models
A small number of inputs in a large dataset can poison training data pretty drastically. Anthropic wrote a good article about it a while back [0]. This should mean its possible to pull back that information fairly easily.
It is hard to not feed it "secrets" too. Models will see path names, read compose files, etc. Of course you can configure things to not leak this type of information, but its not default in most harnesses and isn't 100% sufficient anyways.
If my experience with image generation is any indication, unless AWS keys are somehow extremely prevalent in the training data, you may get something that looks like one, but it definitely won't be valid.
I started using Spark 1.2 for development because if you're willing to let Meta train on your data it was dirt cheap and was actually really pleasantly surprised with it. It's not a frontier model by any means, but for work that didn't require a top of the line model, I really enjoyed using it.
I'm anthropomorphizing it a bit, but it felt like it knew its weaknesses and didn't try to impose it's opinions on me. What I mean by that is that it did what I told it and if there was something unexpected in the code that it put out it was often because I gave it ambiguous or conflicting instructions. It didn't try to go above and beyond and just acted like a tool, which is what I want from a coding agent 90%+ of the time. I also felt that it did a much better job of following established patterns in my code than many of the other current models do. I'm a huge fan of OpenAI's models and Spark 1.2 is what I expected 5.6 Luna to be.
I'm curious and a little excited to use 1.3, but honestly a little worried that as Meta pushes for better benchmarks that Spark will start to fall into the trap of trying to be "helpful" in ways I don't want it to be.
Tangential, but when I first started using Spark 1.2, it made me realize how much I miss 5.3 Codex. That model was the peak of coding models, IMO, in that it knew how to write good code, but didn't try to overstep or be "helpful" in unexpected ways. That got me thinking about how the major labs seem to be stepping away from coding focused models toward more general purpose ones and how I can't help but feel like that's a mistake.
>I started using Spark 1.2 for development because if you're willing to let Meta train on your data it was dirt cheap
its free on opencode and i use it for personal projects. most of my personal projects are AI generated since its personal projects. nothing important are on them. it is hilarious if Meta is training their AI model with AI generated code.
Thats not exactly 'validated'. Feels very noisy, it is not a good bar for either
- does this code do what the user actually asked
- is this code actually 'good'
There would be so many examples of coding projects that these models began or attempted to work in, that were abandoned because the models were floundering.
I would imagine the labs have some decent ways to produce novel requirements and then actually validate they are met, without the noisiness of implicit human feedback.
That said, the more I think about it, you are right, there's probably also very good ways to extract signal for all these sessions.
This is exactly what RLVR is, and the reason that models have improved so much at verifiable domains like coding and math while not so much on unverifiable ones like writing and UI design.
The useful training data is when you clarify your intent, when you tell the model a different approach would be better, when you consistently refactor towards Y and away from X, and so on. The training data isn’t the code, it’s the session transcript. (Anthropic would call this a “distillation attack” against their model, but in this case the model is you!)
I agree that some of the smarter models are actually worse. I hope they take a model that's good enough--there are many--and just try to get it chatjimmy.ai speed.
I have to think that's the future, somehow, and I'm really excited about it.
Would be cool if there was a benchmark to evaluate the “tool-like” quality of a model - its capability to quickly, cheaply, accurately, do exactly as it is asked.
I had no idea Meta has a coding agent harness. Does anyone have experience with it and can comment? The 1.3 contributor prices look very attractive. I'll probably start using their API if performance is good and the API is reliable with decent rate limits.
You should use their harness. They trained it on multiple harnesses but have specifically optimized it for their harness. Cline also did an independent experiment w spark 1.2 where using the native harness makes it use fewer tokens / turns to accomplish tasks
> Co-trained with the harness. Muse Code was in the training loop from day one, so tool calls succeed and plans execute cleanly. Crucially, we trained across multiple harnesses, so while the model is at its best in Muse Code, it still generalizes to other coding agents you already use.
DeepSWE scores 75.4 - that's the best score so far. And it's crazy cheap!
Google held the top a few hours today with Gemini 3.8 Flash, but now second to Spark 1.3. All this competition will drive prices down!
Muse 1.2 wrote a terrible "smart summaries" extension for my pi setup. It was sending every single steamed chunk for summarization instead of waiting for the full CMD.
This is an error I would expect from sonnet 4, not a model that was supposedly just a few points behind sol.
With the contributor pricing being more than 10x cheaper than the standard, that would make it best and cheapest on the DeepSWE leaderboard! It feels fast in my experience too. LLMs keep improving at an insane pace.
and they're ultimately tools strictly to replace you and your labor, they can't/won't cure cancer or make your life better. Your life will get worse and worse in every aspect until they extract maximum value from all of our lives with this technology through every avenue possible. Not sure why you guys are so excited about these developments.
This technology is strictly an extractive parasite on the world. Use it, but don't be excited.
The sibling reply to this is just such lazy thinking, such a trite cliche. Yes, all members of a generation are bad, end of story. Can we get back to the war between the sexes now?
global development and relief of poverty has relied on there being an economic surplus for all from organized labor. everyone gets a benefit although it is unfairly distributed.
i think that there is growing organized labor today that produces no surplus. instead, it transfers wealth from some to others, causing net harm to all in the process. an example of this would be purdue pharma.
depending on who you ask the list of jobs and industries which have zero surplus is getting large. swathes of private equity and leveraged financial instruments, shitcoins, management consultancy, are pure deadweight loss.
+1. I've used the recent Gemini Flash models and I've used Opus 5, and the latter makes the former look like a box of broken crayons. Unless Flash 3.8 and/or this Muse Spark model are a much bigger deal than people seem to think, I will eat my hat if either one can come close to Opus 5 in actual real life "long-horizon software engineering" tasks.
(I'm not happy about the above being true, but it's the reality I seem to inhabit.)
I have been using glm 5.3 flash and it feels as good as opus 5. Put a lot of work into it this week (100m tokens). Now I'm curious to try this one. These smaller models are getting very good imo
A model that (at least in benchmarks) is getting closer to SOTA. A clear separation between what’s used to improve their products and what’s not (at least this is what they claim).
Good job Meta! Seriously. This is almost making me forget about the 18B$ lawsuit for children social media addiction.
It's somewhat useful to note just for your own timelines that Fable was reportedly trained in February. I'm not sure when mythos 5.1 finished training, but muse spark 1.3 almost certainly finished more recently than that.
This doesn't mean it's not one of the best models available (clearly it is), but that table didn't compare Fable/mythos (unless I missed it?) and OpenAI will be releasing a much more recently trained model (Astra) any day.
So you shouldn't think "wow, Facebook has caught up"
You should think, "wow, Facebook is less than 6 months behind the frontier" and that they're actually creating good models which is going to be good in many ways (price for customers, for one!)
There are downsides too, but I'll discuss those separately somewhere
I didn't like 1.2, It make some mistakes in a web app, so I quickly went back to Claude, Kimi K3 or Deepseek V4. Hope this one can clear agentic development, because Muse Spark models are fast and cheap.
Meta is one of those companies where, if there is anything remotely comparable, I'm happy to pay more to not use them. They've had a profoundly negative impact on society and Zuckerberg is not who I want controlling the future at the top of AI.
I feel the same about Grok w/ Elon. I will pay extra to use someone else.
I'm not an Amodei stan, but of all of these people he seems to have the most ethical focus. Again, not everything done perfectly and I have my gripes, but of the leaders of frontier labs, I'll vote with my money.
And, yeah, I wouldn't trust sama to watch my bag while I went to the bathroom.
"Avoid generic tangents" / "Please don't complain about tangential annoyances."
That's pretty much 90% of HN these days.
Apple releases a new iPhone? Here comes the flood of decade-old complaints about long-discontinued Mac butterfly keyboards and walled gardens.
Microsoft releases a new version of Windows? Here come the gripes about Azure.
Google changes something in GMail? Play Store!
It's like there's an army of bots out there determined to reduce the productivity of the Western tech bubble by diverting everyone into endless circular arguments about absolutely nothing of relevance to the topic at hand.
Grandparent comment has zero to do with the article. It's just GP generically bitching about Meta. (Your "quote" of the comment does not appear anywhere in the actual comment.)
If it was up to Dario we'd all be banned from using open-weight models, and we'd have to be investigated for PRC connections before sending our allotted five API queries a week.
SamA better than Zuck? Zuck was at least a kid when he made a lot of his bad decisions, and he seems to be getting much better. Sam is on the reverse trajectory.
Sam is on a delayed trajectory of power, but he surely was not great when he was young either. See: Aaron Swartz calling him a sociopath who could not be trusted, well over a decade ago.
You can ask 100 people and they'll all give you a different list. It's subjective.
I think a less personal ranking would be, as a business owner, which of those providers is more dependable? As in, you don't care about evil, just your stuff working. I think maybe OpenAI?
Google. They have experience operating at scale, and AI is a big enough focus that they won't wind it down. All the big providers are kinda crappy, but if you want reliability, Google is the best option.
Google has experience working for themselves at scale. Your business should never rely on Google more than it is forced to. Even if it's not something they'll wind down, providing acceptable service to anyone is not on their agenda. GCP speaks for itself...
In my personal opinion (this will be controversial and feel free to disagree): Elon is the best.
* great contributions to many industries including spaceflight, electric cars, and self driving cars. It doesn't even matter if he is the technical mind behind these achievements or if he is just a buffoon that pretends to know the implementation details; the dude has a way of bringing together experts, having the overall vision, and managing them properly to ship amazing stuff.
* sane and reasonable takes on AI/LLM stuff. I can't really argue with "pursuit of truth" as the guiding principle. Grok talks normally without "Claudlish", has a balanced score on political bias unlike other models, has a low hallucination rate, is the best at dealing with latest news (unlike ChatGPT that refuses to believe new developments and gaslights the user), and they "never silently downgrade intelligence or fall back to other models."
In contrast, while Dario is doubtless a super smart pioneer in the AI space, his sanctimonious "We know what's good for you" attitude and extreme censorship is really offputting. The lengths to which he tries to ban or hamstring open models seems like an underhanded way to defeat competition. If he were to succeed, it would be a big setback to the thriving ecosystem of open models and hamper the development of the entire industry.
It seems that a rogue engineer poisoned the prompt in that instance. But the fact that they keep the system prompt open is nice. Generally I am biased towards favoring more freedom and openness rather than clamping it down in the name of safety.
They all suck beyond any tolerable threshold. Some of them are further away from the threshold. But at this point, how far each is from the tolerable threshold is besides any point and not worth arguing over. The least of five evils is still evil.
Gotta be honest that I’m tired of the “I hate Zuck and Meta so much” comments every time Meta does anything. Ditto Elon/X. Fine, I get it. I don’t like Zuck either. But the post is about Muse Spark 1.3. What do you think about that? If you don’t like it because Meta made it, then maybe just don’t use it and stay silent.
Technology doesn't just spring into being, there will always be comments on the organizations that developed it. If you don't like them or find them repetitive, it is far easier to collapse them and move on then bend a stranger to your will
I get it, but the underlying problem is: we don't have a society-wide, effective solution to counterbalancing extractive systems. Lacking a reliable label, we have to constantly signal what's on the ingredients list.
Okay, but the comment I reacted to was not that. It was simply (paraphrasing) “I won’t use anything from Zuck/Meta.” If it had been, “Be careful because I have insider information that Zuck/Meta is using Muse Spark to do <insert-nefarious-thing-here>, and here’s my substantiation for that…” I’d be okay with it. That’s interesting information that moves a conversation forward. But it wasn’t. It was just content-free “I don’t like Zuck” nonsense.
What I'm tired of is the top story (or five) on HN every day announcing Spark Opus Fable Grok Gemini v4.1i3-F. Like, who actually cares? Are people excited for the new benchmarks? Is it interesting to read the model cards? And look, part of my job is to use these things and part of my job is to pick EC2 servers, too. The front page of HN is increasingly resembling one of those endless AWS pricing lists.
And yeah, I don't like any of the people or companies building LLMs either. At least the griping is somewhat interesting by comparison. The model isn't news. The news on Hacker News is that other professionals feel the same way.
I think a lot of people are curious where the "knee" is on gains and productivity, particularly in the agentic space, which is where the real value is. A lot of us are being forced to shoe-horn this stuff into existing products, and knowing how much of the task the model can do now, vs having to build a complex custom harness, is valuable information to have. A year and a half ago it took our dev maybe six weeks of struggling with LangChain to approximate what Claude + MCP server can do today. The MCP server took us perhaps 2 days to build and 3 more to get it production ready. Today that MCP server gets 2-3 commits per month. I absolutely want to know when new models come out.
As for smaller models, we run a pretty wide variety of agentic workload doing data enrichment and, increasingly, a bunch of evaluation jobs to alert a human to review certain scenarios etc. These all run on the smaller 27B and 35B class models, and tooling behavior has improved DRAMATICALLY since april. The latest qwen 3.8 model has a 95% success tool call rate during internal testing and about 94% real world. That's about 3% better than the 35B-A3B model we're using today, but the 35B MoE is so much faster then 3% is worth the trade-off.
> Like, who actually cares? Are people excited for the new benchmarks? Is it interesting to read the model cards?
I'm genuinely interested. Even the benchmarks - before Fable came out & while waiting for Astra, I actually setup a math model to predict where they would land (Fable came in at 66 on AA exactly as it predicted), and now I have a model for where these models and Chinese models will likely land in future, and when. And probably no surprise that it's mid-2027 when we cross AA 100, essentially as AI 2027 predicted all along.
I'll probably setup the harness I made for myself to try out some of these models on OpenRouter. I've been frustrated with Opus & Fable 5 and found that I like working with GLM 5.3 Flash far more than I expected to, and I only found that out because I tried it during the stealth Ox Alpha launch, which I probably found out about here too.
TLDR, I think some / many people here are genuinely interested, excited, and that's why they're upvoted so highly. And Muse Spark 1.3 scoring highly seems like a genuine surprise, when Meta was basically a write-off not long ago.
Okay but Muse Glimmer 30B is one of the best small open weight models today, and IMO the best from a US lab (only real comparison is Gemma4 dense right now).
By their own benchmarks it is about 10% lower scoring than Qwen 3.6 35b-a3b, but I've added it to my list. Always looking for MoE to compare to it so we can squeeze more out of our local LLM system.
Yeah this would be a great point if it were true and they didn’t give Mythos access to companies to fix bugs, which they did and have.
It’s genuinely a difficult question. Not black and white. The models are really good at finding bugs, as demonstrated by people using Fable to reverse engineer. People make it sound like he’s just making it up.
This would be more convincing if mythos was something uniquely special and not something merely a couple months ahead of everyone else. It was great marketing though.
They gave access, but considering that they wouldn't even sign the "don't ban open weights" letter, it's clear they would prefer to have tight control over who they bless with that access.
I'm the guy you replied to, apologies for using a different account I'm away from my computer now.
The distinction to me is that Anthropic gives access to that model but doesn't give control. They reserve the right to cut you off if they don't like what you are doing and require you allow data retention for Fable and Mythos to ensure your are not up to any skullduggery.
Meta, Alibaba, Mistral, even OpenAI has released models users can run locally and fully control. That is a whole world of difference.
Dario's "ethical" look is also kinda sus. I hate to use ad hominem, but the dude's wife literally pitched a porn film to Epstein even after he was a convicted registered sex offender [1]. Dario is also really sinophobic (it is commonly claimed in Chinese AI circles that his former employment at Baidu triggered him so much that he harbors a personal grudge against the entire race).
> I'm not an Amodei stan, but of all of these people he seems to have the most ethical focus. Again, not everything done perfectly and I have my gripes, but of the leaders of frontier labs, I'll vote with my money.
Amodei is NO Saint!!! He's the most savvy in drumming up the AI doomsday scenarios and haven't yet to apologized his failed forecast of Claude taking over 90% of the coding jobs.
Funny. Dario seems like the biggest snake in the industry to me and has leaned the hardest into doom marketing out of all of the influential leaders. With Altman (or Google), it's a transaction, and that's something I can live with.
I just don’t see how people have looked at what has happened with Mythos and the deluge of fixes from companies, then come to this conclusion.
He has a really hard job. He errs on the side of conservatism in releasing and then people get Really Mad.
Safeguards on cybersecurity are not great for Anthropic revenue! As evidenced by people getting pissed, moving to Sol, and them having a smaller market for what Fable can do.
It’s clearly bad for revenue and not great advertising to say, “you can’t use this but here is a nerfed version that will annoy you and not solve important problems.”
Anthropic/Amodei have been the most alarmist about model safety, so multiple things can be true. A lot of tech companies avoided scrutiny by sending bribes to Trump (naked corruption is bad, I'd rather nobody do that), Anthropic didn't...so, combined with their fear-mongering about the danger of Mythos and open models (which seems aimed at regulatory capture) and the lack of bribes flowing to the Trump administration, they got stepped on by the federal government based on the excuse Anthropic provided.
I dunno. Everybody seems to be playing pretty dirty. Some people have a much longer history of that, though. Obviously, Meta and Musk are outliers even in an industry full of problematic behavior.
Security vulnerability capability is not the only thing they're scare-mongering about. They're the biggest purveyors of the, "We think the little guy in the computer who is made of algebra might be a real live boy and he might want to kill all of humanity when he grows up," line of alarmism.
That's ok to think at this point, given the trajectory of the last few years. Certainly it's one of those things where erring (marginally and slightly) on the side of being safe about it is better than the alternative.
And he drew a red line wrt the Pentagon's use of Anthropic's models for autonomous weapons and surveillance of American citizens, and he stood by it, even when the government took steps to materially damage the company. This required true courage. Name me another CEO, of any major American company, that has demonstrated this much fortitude.
Meta and Microsoft are two of the absolute worst evil companies on earth and Amodei is trying very hard to join them.
These Effective Altruists are despicable people: a bunch of thieves working to line up their own pockets while posturing as a force of good.
Remember that they schemed to not only present SBF as the 2nd coming of Christ (including in the NYT and in Forbes) but to also give him a voice after his scam had been uncovered. Thankfully, the judge didn't have any of this Effective Altruist bullshit.
SBF invested 500 millions of misappropriated funds in his buddy from the EA movement's Anthropic company (and, thankfully, the judge forced those shares to be sold: so SBF didn't get to be a billionaire).
You cannot hate enough people who say that harming others for the greater good is justified.
Then of course, already mentioned in this thread, there's the whole Epstein/Amodei's "I'm in the porn business" wife connection (where you don't need to squint much to see young women abused).
These kind of people are the absolute worst scum on this earth.
Strong disagree with the Anthropic being good at all part. This is not defending anyone else, but…
Anthropic leadership repeatedly presents themselves as uniquely morally qualified to steward agi and decide how humanity should get access to it. Yet they have repeatedly failed basic morality tests.
Pirating books for financial gain. The newer Sony/Warner music case shows this is pattern behavior.
Aggressively scraping other people's works, despite the authors' requests not to do so.
Then applying massive usage restrictions on their own work.
And probably the most disqualifying is backing away from their own hard AI safety commitments.
It's almost like running a trillion-dollar business with neck-to-neck competition against other frontier labs and even state-sponsored efforts requires some ethical trade-off.
Pirating books is just straight up morally correct. I don't like Anthropic's bullshit "safety" filters, but training on shadow library data? Yeah no, it makes sense.
It makes a lot more sense than having to work around copyright by scanning out physical books. Unfortunately, one was ruled legal and the other was not.
Which safety commitments did they back away from? My understanding is that they believe safety can only be researched from the frontier, and so they're trying to be pragmatic to stay near the frontier (and viable) in their choices.
From what I know, the "books3" dataset was normalised in the LLM and research ecosystem, where collected datasets were seen as valid to train on and/or fair use. I'm not sure any of the major frontier companies are free from that, if we don't believe it was fair use.
I do think most of their choices are explainable by "they just believe in agi risk". You truly wouldn't want non-agi-pilled companies to train on your data and approach the frontier if you were worried. You might slightly hurt your own business with safety filters (that no one else does) if you were worried. They are less worried about other "moral" decisions like "sharing" if they conflict with AGI: the research they still share is all of their safety research.
This definitely doesn't make them "good", but they do seem fairly "consistent". Most of these issues were talked about publicly by the founders long before Anthropic was founded and/or the AI race+money appeared.
as a safety commitment they walked away from - they were similarly negligent to openai in terms of asking a model with a hacking based harness to go have fun, and then not watching it at all while it could do harmful and illegal stuff.
thats not something you expect from a company that "believes in agi risk"
I don't think "not watching it at all" is completely fair. They thought they had sandboxing/monitoring etc. I definitely won't say they're free of mistakes though.
Note that the companies that haven't faced these issues so far are the ones that don't do safety testing, or don't have frontier models. I'm not sure who I would pick as "better" on any of this right now.
I will give Anthropic credit for standing up against the department of war. The bar is incredibly low, but not doing domestic surveillance and not creating autonomous weapons are laudable.
That doesn’t mean I like them pirating books and being shady about tokens and paternalistic “safety”
It's hard for me to see much difference between Amodei and Sama. My guess is they're both savvy SV CEOs who will bend their message, alliances and principles pretty far if that's what it takes to get ahead. Musk and Zuck feel like something else entirely, with all the reactionary imagery, populist bullshit and the societal damage around their platforms.
He wants to build a tech-god kept in chains whose power he parcels out to the unwashed masses he deems worthy like some sort of high priest of intelligence.
And that is being charitable and going by the interpretation that he actually believes what he says.
Well what do you want? Presenting clear, desirable, and achievable visions and trying to build consensus for how AI should develop is crucial at this point in time.
I hate Meta main business, but you have to admit that on the non business related and open source side, they have released amazing things that changed the world.
React for example.
And we could easily guess that there wouldn't have been so much open source models, and grand public experiments and free tools if llama models were not release to the general public.
Anthropic is not exactly a saint either. I had a recent issue where they denied fable credits even though I was hospitalized during the claim period. I have annual plan with them.
As much as everyone hates sama, I think OpenAI is much more of a company with good marketing and sales team.
I'll happily pay for Grok, it's a great model. 4.6 often does better than Anthropic at coding and analysis where Anthropic fails for 'oh no cyber security, don't ask me to check if you're redacting passwords correctly in logs'. And it has no problem telling the truth where OpenAI / Anthropic don't want to upset the people on the left and will happily lie or avoid hard truths.
Edit: I get it. It's a hard pill to swallow. I understand people don't like Musk or Zuck. But it doesn't change the fact that you're being lied to and brainwashed.
Muse Spark may be competitive in capabilities but it’s not for serious works since Meta trains on your prompts so no ZDR, in contrast Chinese provider like Z.AI promises ZDR which is more attractive to big corps.
Meta also has 50k engineers. Not to mention that tons of meta infrastructure - including ads! - use AI. Would you want that sort of business be this dependent on someone else?
All people here care about is hating Meta. Just look at the top voted comment. No one cares about the merits of the model, etc. HN has become nothing but an echo chamber.
muse-spark-1.3-contributor. Say what you want and Meta, changing the pricing to explicitly say 'we train on this and value it this much' is what every model provider should do. As a side note, it is now completely obvious how much stealing my tokens for training is worth to model providers. I avoid/pay extra/try my best to make sure I am not getting trained on but it seems like it keeps popping up that I missed a setting somewhere. This is the first quantifiable number I have seen out there from a model provider. Maybe it can help in lawsuits to quantify the damages for copyright/other things?
This has been my hunch for a while about all the discourse of "OpenAI/Anthropic subscription pricing is unsustainable!!"
We understand theoretically they're taking our data, but yeah, that data is vital to the entire business plan of all these companies and WAY more valuable than people are giving credit for.
I checked up on Mistral recently and saw their Claude-alike coding harness is using GLM now, whatever it takes to keep users on their platform and feeding them data.
It's a toggle. Some will automatically enable it and you have to turn it off. People who rapidly click through setup flows can miss it and leave it enabled.
I have not tried Muse Spark for code, but I've been using it for a while to write Latin. I find it's one of the best at it, alongside Gemini. For example, I've recently been using it to translate the subtitles of the show I'm watching into Latin, to provide me with a bit more input. (I'm learning Latin, for context)
Very keen to try this after using Claude Code over the last few months.
Should I just point Claude Code to Muse Spark endpoint (because I'm familiar with Code)? What do people think of Muse Code or other coding agent harnesses?
Well thats very interesting. Thank you.
Will be interesting to see how hard/easy it is to translate my Claude skills, loop design, etc to the new harness.
This kind of raises another question to me regarding the coding benchmarks, how much of it is model versus harness?
You can check out Muse Spark 1.3 by using OpenCode (https://opencode.ai/ - open-source AI / coding harness). There's a terminal version and a GUI / desktop version. Good luck!
I am very impressed by this model so far. It's faaast and it seems to be just intelligent enough to do really well. It's UI work (simple python UI) is very clean and functional. The UX was 'there'.
$META has everything it needs, great team, great models coming out, great infrastructure (GPUs), great userbase and distribution channels. $META is underrated.
it seems like gemini 3.8 flash is more capable and cheaper. The only reason i would use this is if i was willing to share my data with meta, and allow them to train on my data. In that case it becomes dirt cheap.
For folks who are impressed with costs, why does it matter to you? Is subscriptions not a thing? I may be missing something but only companies should really care about this I would think?
Im a caveman writing c/cpp. Last time ms1.2 was even worth than DeepSeek v4f preview on internal benchmark. It just feels like extremely over fitting on certain paths.
frozenseven | 9 hours ago
https://news.ycombinator.com/item?id=49541149
simonw | 9 hours ago
4.2266 cents, 38 seconds.
For comparison here's Muse Spark 1.2, which animated it without me asking it to: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
The 1.3 one is definitely better - better bicycle frame, better wing, better pelican hat.
UPDATE: Here's another one with five pelicans for each of the five Muse Spark 1.3 reasoning levels: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
The most expensive was reasoning level xhigh - 7.5 cents, 1m34s.
And I ran five pelicans at all reasoning levels for 1.2 as well, here: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
jmkni | 9 hours ago
Definitely an upgrade over 1.2
drusepth | 9 hours ago
simonw | 9 hours ago
The 2D / flat ground feels reasonable for a SVG, which implies a vector illustration.
piker | 8 hours ago
m12k | 8 hours ago
johntb86 | 8 hours ago
vunderba | 9 hours ago
It's like when you ask your average person off the street to draw a house - it'll almost always be square with a triangle roof, one door, and two windows.
In the pelican/bike example, it's probably a bit of a self-perpetuating snowball too. If the earliest examples were bike left-to-right, flat ground, etc. then they are also being scraped up in future LLMs.
polyterative | 8 hours ago
srcreigh | 8 hours ago
https://www.ikea.com/ca/en/p/barndroem-box-beige-70560615/
https://www.ikea.com/ca/en/p/vallaby-rug-green-10548216/
collabs | 8 hours ago
There is this scene in the HBO series Westworld where a "host" says some words in sequence which is shown on a display as she says it. Of course, even me thinking of this scene and connecting it to your comment was not original, someone else clearly had the same programming as me.
A medium blog post says
> Pair what with me?” — the moment Maeve (a humanoid android) uttered those words in Westworld (Season 1, Episode 6: “The Adversary”), something clicked. Not for the average viewer, but for me, a STEM educator and AI enthusiast who, just weeks earlier, had read Stephen Wolfram’s seminal essay, What Is ChatGPT Doing … and Why Does It Work?
ACCount37 | 5 hours ago
It's not even that old - but back when it was aired, an AI that can not just string together coherent sentences, but produce coherent reactions in novel, fully unintended contexts, like Maeve was doing there? It was totally a sci-fi premise.
Now we have AIs capable of that and more, and no one bats an eye.
twoodfin | 2 hours ago
Required sci-fi suspension-of-disbelief in 2017, and then at some point in the last few years we just blew by that one.
Later seasons of the show were much less dramatically satisfying, but also played out the consequences of the science of artificial intelligence demonstrating as a side-effect that human intelligence and free will might have as much of an uncertain foundation as that of machines.
How much data from the Panopticon, how many parameters would it take to train a model that could predict your responses?
sly010 | 25 minutes ago
irthomasthomas | 3 hours ago
Gigachad | 5 hours ago
werdnapk | 5 hours ago
postalcoder | 9 hours ago
daemonologist | 8 hours ago
(Why the drivetrain is on the right, I don't know. But most bike parts follow open standards so it's quite entrenched.)
georgemcbay | 8 hours ago
While I'm sure this factors into things for advertisements for bike components, there is also just a general preference that westerners have for left-to-right motion. Not just in bike ads, but all ads with (or suggesting) movement. And also not just ads, but movies where directors believe left-to-right motion is associated with progression and right-to-left motion is regressive.
kibae | 8 hours ago
labcomputer | 5 hours ago
Bicycle frames are not fully symmetric left-right because you need things like a mount point for the derailleur hanger, and optionally affordances to keep the chain off the stays when the wheel is removed.
Those things have to be on the same side as the chain. Bikes designed for disc brakes additionally need a mount point for the brake caliper on the opposite side from the chain.
Additionally, rear wheels are not symmetric: the spokes on the chain side connect to the hub closer to the plane of the rim. That is, they are more perpendicular to the wheel’s rotational axis than spokes on the opposite side (which is why you should always mount a single pannier on the chain side). This asymmetry is to provide space for the gears.
So once the industry decided to put the chain on the ride, you can’t very well make a group set designed for a left chain if you want it to work on the vast majority of frames.
threetonesun | 8 hours ago
ModernMech | 8 hours ago
optimalsolver | 8 hours ago
reaperducer | 8 hours ago
Because they're computers. They don't have an imagination and the ability to create things from whole cloth the way humans do.
Much like a mother pelican, they regurgitate what they've been fed.
BeetleB | 7 hours ago
https://blog.nawaz.org/posts/2025/Oct/pelican-on-a-bike-rayt...
I plan to update it with more pelicans from all the models released since.
(Spoiler alert: They haven't improved much since then).
xhrpost | 6 hours ago
murkt | 2 hours ago
fc417fc802 | 2 hours ago
> GPT-5.1 Codex
> monstrosity
What are you talking about? That's clearly a sci-fi pelican on a hoverboard (successor of the humble bicycle) wearing a visor. Truly visionary.
porphyra | 6 hours ago
__MatrixMan__ | 6 hours ago
I bet if it instead had something to do with black widow spiders we'd find that we're most often looking at the bottom of the spider's abdomen, regardless of whatever non-spider-like activity is supplied.
SV_BubbleTime | 3 hours ago
They’re all so close in proportions.
elfly | 3 hours ago
bodeadly | an hour ago
tomrod | 9 hours ago
Also 3X token use vs. 1.2
_puk | 8 hours ago
jonahx | 9 hours ago
Fergusonb | 8 hours ago
jonplackett | 8 hours ago
jttnr | 8 hours ago
EugeneOZ | 8 hours ago
Thank you for doing this, I love your benchmark the most!
drob518 | 8 hours ago
gpt5 | 4 hours ago
hollowturtle | 7 hours ago
tintor | 7 hours ago
nojs | 6 hours ago
It’s not deliberate “benchmaxxing” but things that are discussed a lot online are naturally things that LLMs learn better.
fc417fc802 | 2 hours ago
BeetleB | 7 hours ago
https://news.ycombinator.com/item?id=49538333
cheesecakegood | 4 hours ago
tintor | 7 hours ago
Did any LLM draw the front bicycle wheel correctly? ie. center of front wheel slightly AHEAD of steering wheel axis. This is done for bicycle stability.
TiredOfLife | 7 hours ago
Bird knees bend same way human ones do
0xbadcafebee | 7 hours ago
ipsum2 | 7 hours ago
leumon | 6 hours ago
wewewedxfgdf | 6 hours ago
Aced it, got the job as a senior software engineer.
The interviewers afterwards said "it is SO refreshing to find a software developer who actually knows how to code - never seen such a high performance focused, well built pelican on a bike - you have the skills we need".
fuddle | 5 hours ago
sroussey | 4 hours ago
labrador | 3 hours ago
hunterpayne | 3 hours ago
rattray | 3 hours ago
bensyverson | 3 hours ago
nycdatasci | 3 hours ago
smashah | 3 hours ago
latentsea | 3 hours ago
keeda | 37 minutes ago
UltraSane | 3 hours ago
treebeard901 | an hour ago
salutis | 2 hours ago
That is the best joke I have heard this year. Ready for a stand-up comedy special. Or a song. Superb!
pavs | 5 hours ago
I am guessing its not super common, but it happens just so you know.
andytratt | 4 hours ago
m00dy | 2 hours ago
simonw | 17 minutes ago
dwaite | 34 minutes ago
dhon_ | 32 minutes ago
Gecko4072 | 9 hours ago
WASDx | 9 hours ago
scotty79 | 9 hours ago
dominotw | 9 hours ago
redox99 | 9 hours ago
samuelknight | 9 hours ago
schopra909 | 9 hours ago
Moreover,I think it’s impossible to know if you’re hitting a portion of the sigmoid, because there will often be an idea that changes the trajectory altogether.
In 2024, there was a ton of talk about the plateau. Reasoning was an iteration on chain of thought, but it didn’t really work. Deepseek proposes RLVR as a way to get around the lack of $ they have to produce human reasoning trace data. That small iteration catches the eye of OpenAI and Anthropic, turns out to be way more important than even DeepSeek could have ever expected when it comes to improving LLMs for coding, and last 18 months have been an exercise on riding that insight to the nth degree.
That one small iteration brought us a lot of progress. Now we’re seemingly exhausting the impact of that one insight, but there may be another soon enough.
stymaar | 8 hours ago
What was the difference between what deepseek did for R1 and what OpenAI did for o1?
npn | an hour ago
refulgentis | 8 hours ago
Philpax | 3 hours ago
schopra909 | an hour ago
Since DeepSeeks GRPO, they’ve been improvements as well like AliBabas GSPO that have gotten wide adoption. Again iterations
ipsum2 | 7 hours ago
danielmarkbruce | 6 hours ago
gdiamos | 5 hours ago
wxw | 9 hours ago
Definitely shows how important a user data flywheel is for RL and model improvement.
finnjohnsen2 | 9 hours ago
Given this is Meta, my immediate assumptions that one is cheap because it lets me "be the product". I know I'm rushing to conclusions but there is zero trust here. The brain will do its thing. And the wording here is giving the brains a lot of wiggle room.
whimsicalism | 9 hours ago
thefreeman | 9 hours ago
bigyabai | 8 hours ago
Jcampuzano2 | 8 hours ago
I don't see the wiggle room at all.
duplessitous | 8 hours ago
IshKebab | 8 hours ago
In other words, it's not that Meta really wants your data and they're willing to pay top dollar for it. It's that companies really don't want Meta to have their data and they're willing to pay top dollar for that.
zhoBEENG | 7 hours ago
K0balt | 9 minutes ago
warkdarrior | 7 hours ago
r_lee | 3 hours ago
majerep | 9 hours ago
mgaunard | 9 hours ago
tonyhart7 | 8 hours ago
geooff_ | 9 hours ago
oofbey | 5 hours ago
HDBaseT | 5 hours ago
This price/intelligence beats even legacy DeepSeek V4 Flash pricing.
[1] https://artificialanalysis.ai/#total-cost-tabs
meerita | 9 hours ago
7734128 | 9 hours ago
0xbadcafebee | 7 hours ago
HDBaseT | 5 hours ago
It is hard to not feed it "secrets" too. Models will see path names, read compose files, etc. Of course you can configure things to not leak this type of information, but its not default in most harnesses and isn't 100% sufficient anyways.
[0] https://www.anthropic.com/research/small-samples-poison
owaiswiz | 5 hours ago
HDBaseT | 2 hours ago
userbinator | an hour ago
wrsh07 | 5 hours ago
superfrank | 8 hours ago
I'm anthropomorphizing it a bit, but it felt like it knew its weaknesses and didn't try to impose it's opinions on me. What I mean by that is that it did what I told it and if there was something unexpected in the code that it put out it was often because I gave it ambiguous or conflicting instructions. It didn't try to go above and beyond and just acted like a tool, which is what I want from a coding agent 90%+ of the time. I also felt that it did a much better job of following established patterns in my code than many of the other current models do. I'm a huge fan of OpenAI's models and Spark 1.2 is what I expected 5.6 Luna to be.
I'm curious and a little excited to use 1.3, but honestly a little worried that as Meta pushes for better benchmarks that Spark will start to fall into the trap of trying to be "helpful" in ways I don't want it to be.
Tangential, but when I first started using Spark 1.2, it made me realize how much I miss 5.3 Codex. That model was the peak of coding models, IMO, in that it knew how to write good code, but didn't try to overstep or be "helpful" in unexpected ways. That got me thinking about how the major labs seem to be stepping away from coding focused models toward more general purpose ones and how I can't help but feel like that's a mistake.
MangoCoffee | 8 hours ago
its free on opencode and i use it for personal projects. most of my personal projects are AI generated since its personal projects. nothing important are on them. it is hilarious if Meta is training their AI model with AI generated code.
KptMarchewa | 7 hours ago
TiredOfLife | 7 hours ago
dakolli | 6 hours ago
dcl | 5 hours ago
dakolli | 2 hours ago
dcl | an hour ago
There would be so many examples of coding projects that these models began or attempted to work in, that were abandoned because the models were floundering.
I would imagine the labs have some decent ways to produce novel requirements and then actually validate they are met, without the noisiness of implicit human feedback.
That said, the more I think about it, you are right, there's probably also very good ways to extract signal for all these sessions.
avarun | an hour ago
superfrank | 5 hours ago
CGamesPlay | 4 hours ago
sejje | 7 hours ago
I agree that some of the smarter models are actually worse. I hope they take a model that's good enough--there are many--and just try to get it chatjimmy.ai speed.
I have to think that's the future, somehow, and I'm really excited about it.
monkpit | 44 minutes ago
tinyhouse | 8 hours ago
meric_ | 7 hours ago
dcl | 4 hours ago
meric_ | 3 hours ago
Muse code: https://developer.meta.com/ai/resources/blog/build-with-muse...
> Co-trained with the harness. Muse Code was in the training loop from day one, so tool calls succeed and plans execute cleanly. Crucially, we trained across multiple harnesses, so while the model is at its best in Muse Code, it still generalizes to other coding agents you already use.
dcl | an hour ago
tinyhouse | an hour ago
bertili | 8 hours ago
cbg0 | 8 hours ago
bermudi | 8 hours ago
This is an error I would expect from sonnet 4, not a model that was supposedly just a few points behind sol.
gpt5 | 4 hours ago
dominotw | 8 hours ago
WASDx | 8 hours ago
dakolli | 6 hours ago
This technology is strictly an extractive parasite on the world. Use it, but don't be excited.
skybrian | 6 hours ago
switchbak | 3 hours ago
comicjk | 6 hours ago
dakolli | 5 hours ago
atemerev | 5 hours ago
Well, that's about the same validity as "In Western astrology..." or "in flat earth theory..."
monkpit | 39 minutes ago
lukewarm707 | 2 hours ago
i think that there is growing organized labor today that produces no surplus. instead, it transfers wealth from some to others, causing net harm to all in the process. an example of this would be purdue pharma.
depending on who you ask the list of jobs and industries which have zero surplus is getting large. swathes of private equity and leveraged financial instruments, shitcoins, management consultancy, are pure deadweight loss.
the work does nothing or causes net harm.
monkpit | 40 minutes ago
nl | 5 hours ago
That's the opposite of parasitic.
switchbak | 3 hours ago
cycrutchfield | 2 hours ago
yipinwong | an hour ago
People already started using contributor API, and your input is irrelevant.
notatoad | 3 hours ago
anybody who's used these models knows that their real-world software engineering performance has no relation to the ranking on deepSWE.
caconym_ | 2 hours ago
(I'm not happy about the above being true, but it's the reality I seem to inhabit.)
jdm2212 | an hour ago
zackify | 44 minutes ago
israrkhan | 3 hours ago
Compare that to Muse spark 1.3
$1.25/M input, $4.25/M output (without data sharing) $0.10/M input, $0.20/M output (with data sharing)
It is dirt cheap, but only if you are willing to share your data with meta and allow them to use it for improving their models and products.
Lucasoato | 8 hours ago
Good job Meta! Seriously. This is almost making me forget about the 18B$ lawsuit for children social media addiction.
dbbk | 8 hours ago
ctolsen | 7 hours ago
pqdbr | 6 hours ago
cdelsolar | 16 minutes ago
neuronic | 6 hours ago
switchbak | 2 hours ago
wrsh07 | 5 hours ago
This doesn't mean it's not one of the best models available (clearly it is), but that table didn't compare Fable/mythos (unless I missed it?) and OpenAI will be releasing a much more recently trained model (Astra) any day.
So you shouldn't think "wow, Facebook has caught up"
You should think, "wow, Facebook is less than 6 months behind the frontier" and that they're actually creating good models which is going to be good in many ways (price for customers, for one!)
There are downsides too, but I'll discuss those separately somewhere
ChrisArchitect | 8 hours ago
sunaookami | 8 hours ago
Lmao. And their benchmark table only shows max reasoning.
mromanuk | 8 hours ago
tyre | 8 hours ago
I feel the same about Grok w/ Elon. I will pay extra to use someone else.
I'm not an Amodei stan, but of all of these people he seems to have the most ethical focus. Again, not everything done perfectly and I have my gripes, but of the leaders of frontier labs, I'll vote with my money.
And, yeah, I wouldn't trust sama to watch my bag while I went to the bathroom.
loeg | 8 hours ago
reaperducer | 8 hours ago
That's pretty much 90% of HN these days.
Apple releases a new iPhone? Here comes the flood of decade-old complaints about long-discontinued Mac butterfly keyboards and walled gardens.
Microsoft releases a new version of Windows? Here come the gripes about Azure.
Google changes something in GMail? Play Store!
It's like there's an army of bots out there determined to reduce the productivity of the Western tech bubble by diverting everyone into endless circular arguments about absolutely nothing of relevance to the topic at hand.
_diyar | 8 hours ago
> Meta announces they have a new model, demonstrating its capabilities.
> Parent comment states „regardless of this model‘s specific capabilities, if I can avoid it I will.“
loeg | 8 hours ago
optimalsolver | 8 hours ago
redox99 | 8 hours ago
They all suck. Pick your poison.
kenjackson | 8 hours ago
Here's the order, from best to worst.
Amodei
Google
SamA
Zuck
Elon
scottyah | 8 hours ago
Zambyte | 8 hours ago
KptMarchewa | 7 hours ago
>>> They "trust me" >>> Dumb fucks
redox99 | 8 hours ago
I think a less personal ranking would be, as a business owner, which of those providers is more dependable? As in, you don't care about evil, just your stuff working. I think maybe OpenAI?
a2ff6eeb0 | 8 hours ago
applfanboysbgon | 8 hours ago
ralusek | 8 hours ago
a2ff6eeb0 | 7 hours ago
xnx | 6 hours ago
redox99 | 8 hours ago
utopcell | 8 hours ago
True. With 5 choices you need at least 126 people before you can guarantee that two lists are the same.
porphyra | 7 hours ago
* great contributions to many industries including spaceflight, electric cars, and self driving cars. It doesn't even matter if he is the technical mind behind these achievements or if he is just a buffoon that pretends to know the implementation details; the dude has a way of bringing together experts, having the overall vision, and managing them properly to ship amazing stuff.
* sane and reasonable takes on AI/LLM stuff. I can't really argue with "pursuit of truth" as the guiding principle. Grok talks normally without "Claudlish", has a balanced score on political bias unlike other models, has a low hallucination rate, is the best at dealing with latest news (unlike ChatGPT that refuses to believe new developments and gaslights the user), and they "never silently downgrade intelligence or fall back to other models."
In contrast, while Dario is doubtless a super smart pioneer in the AI space, his sanctimonious "We know what's good for you" attitude and extreme censorship is really offputting. The lengths to which he tries to ban or hamstring open models seems like an underhanded way to defeat competition. If he were to succeed, it would be a big setback to the thriving ecosystem of open models and hamper the development of the entire industry.
billypilgrim | 6 hours ago
porphyra | 6 hours ago
alex1138 | an hour ago
cactca | 7 hours ago
runarberg | 6 hours ago
drob518 | 8 hours ago
idiotsecant | 8 hours ago
duplessitous | 8 hours ago
canadaduane | 8 hours ago
drob518 | 8 hours ago
luckylion | 8 hours ago
That's obviously not the issue with that -- you don't see those comments on Google's AI announcements.
owebmaster | 7 hours ago
jesse_dot_id | 8 hours ago
troupo | 8 hours ago
That:
- like all models it was trained on stolen data
- additionally it was trained on Facebook users who were all opted in to AI training with a convoluted 10+ step process to opt-out of
> If you don’t like it because Meta made it, then maybe just don’t use it and stay silent.
Why should anyone stay silent?
monster_truck | 8 hours ago
whateveracct | 8 hours ago
Anduril makes this same complaint whenever their job posts get dumped on. Same idea. Fix your bad PR, buddies :)
noduerme | 8 hours ago
And yeah, I don't like any of the people or companies building LLMs either. At least the griping is somewhat interesting by comparison. The model isn't news. The news on Hacker News is that other professionals feel the same way.
hadlock | 7 hours ago
As for smaller models, we run a pretty wide variety of agentic workload doing data enrichment and, increasingly, a bunch of evaluation jobs to alert a human to review certain scenarios etc. These all run on the smaller 27B and 35B class models, and tooling behavior has improved DRAMATICALLY since april. The latest qwen 3.8 model has a 95% success tool call rate during internal testing and about 94% real world. That's about 3% better than the 35B-A3B model we're using today, but the 35B MoE is so much faster then 3% is worth the trade-off.
NamlchakKhandro | 6 hours ago
SyneRyder | 6 hours ago
I'm genuinely interested. Even the benchmarks - before Fable came out & while waiting for Astra, I actually setup a math model to predict where they would land (Fable came in at 66 on AA exactly as it predicted), and now I have a model for where these models and Chinese models will likely land in future, and when. And probably no surprise that it's mid-2027 when we cross AA 100, essentially as AI 2027 predicted all along.
I'll probably setup the harness I made for myself to try out some of these models on OpenRouter. I've been frustrated with Opus & Fable 5 and found that I like working with GLM 5.3 Flash far more than I expected to, and I only found that out because I tried it during the stealth Ox Alpha launch, which I probably found out about here too.
TLDR, I think some / many people here are genuinely interested, excited, and that's why they're upvoted so highly. And Muse Spark 1.3 scoring highly seems like a genuine surprise, when Meta was basically a write-off not long ago.
abjhn | 5 hours ago
dangoljames | 7 hours ago
georgespencer | 7 hours ago
You might consider following your own advice.
fouc | 8 hours ago
Yajirobe | 8 hours ago
aftbit | 8 hours ago
Bluestein | 7 hours ago
hadlock | 6 hours ago
tyre | 7 hours ago
HDBaseT | 6 hours ago
The problem inference providers will not be able to get anywhere near the contributor pricing.
jwitthuhn | 7 hours ago
tehlike | 7 hours ago
tyre | 7 hours ago
It’s genuinely a difficult question. Not black and white. The models are really good at finding bugs, as demonstrated by people using Fable to reverse engineer. People make it sound like he’s just making it up.
phoghed | 7 hours ago
hgoel | 7 hours ago
throwaway63486 | 7 hours ago
The distinction to me is that Anthropic gives access to that model but doesn't give control. They reserve the right to cut you off if they don't like what you are doing and require you allow data retention for Fable and Mythos to ensure your are not up to any skullduggery.
Meta, Alibaba, Mistral, even OpenAI has released models users can run locally and fully control. That is a whole world of difference.
codexon | 5 hours ago
Half a year later, it is still not available to everyone else.
porphyra | 7 hours ago
[1] https://www.forbes.com/sites/alisondurkee/2026/08/14/who-is-...
devy | 7 hours ago
Amodei is NO Saint!!! He's the most savvy in drumming up the AI doomsday scenarios and haven't yet to apologized his failed forecast of Claude taking over 90% of the coding jobs.
samtheprogram | 7 hours ago
bradlys | 7 hours ago
jansport123 | 6 hours ago
platinumrad | 7 hours ago
tyre | 7 hours ago
He has a really hard job. He errs on the side of conservatism in releasing and then people get Really Mad.
Safeguards on cybersecurity are not great for Anthropic revenue! As evidenced by people getting pissed, moving to Sol, and them having a smaller market for what Fable can do.
It’s clearly bad for revenue and not great advertising to say, “you can’t use this but here is a nerfed version that will annoy you and not solve important problems.”
SwellJoe | 7 hours ago
I dunno. Everybody seems to be playing pretty dirty. Some people have a much longer history of that, though. Obviously, Meta and Musk are outliers even in an industry full of problematic behavior.
felixgallo | 7 hours ago
SwellJoe | 4 hours ago
felixgallo | 3 hours ago
adriand | 6 hours ago
codexon | 5 hours ago
TacticalCoder | 7 hours ago
These Effective Altruists are despicable people: a bunch of thieves working to line up their own pockets while posturing as a force of good.
Remember that they schemed to not only present SBF as the 2nd coming of Christ (including in the NYT and in Forbes) but to also give him a voice after his scam had been uncovered. Thankfully, the judge didn't have any of this Effective Altruist bullshit.
SBF invested 500 millions of misappropriated funds in his buddy from the EA movement's Anthropic company (and, thankfully, the judge forced those shares to be sold: so SBF didn't get to be a billionaire).
You cannot hate enough people who say that harming others for the greater good is justified.
Then of course, already mentioned in this thread, there's the whole Epstein/Amodei's "I'm in the porn business" wife connection (where you don't need to squint much to see young women abused).
These kind of people are the absolute worst scum on this earth.
ls_stats | 7 hours ago
biddit | 7 hours ago
Anthropic leadership repeatedly presents themselves as uniquely morally qualified to steward agi and decide how humanity should get access to it. Yet they have repeatedly failed basic morality tests.
Pirating books for financial gain. The newer Sony/Warner music case shows this is pattern behavior.
Aggressively scraping other people's works, despite the authors' requests not to do so.
Then applying massive usage restrictions on their own work.
And probably the most disqualifying is backing away from their own hard AI safety commitments.
nostromo | 7 hours ago
They want to position AI as an insurmountable threat in order to regulate away any future competitors. They’re trying to speedrun regulatory capture.
sscaryterry | 7 hours ago
metadat | 7 hours ago
Humans naturally want SOMEONE to be the good guy! Sad story, in this instance.
dofm | 7 hours ago
Which one? The main bit that reports to Daniela Amodei, or the little comfort blanket cabinet around Dario and his "chief of staff"?
There is a leadership branch that can pretend to be morally qualified and aware and to think about the big picture and ethics.
It is at least somewhat remote from the bit that is doing the actual business things.
vovavili | 6 hours ago
ACCount37 | 6 hours ago
It makes a lot more sense than having to work around copyright by scanning out physical books. Unfortunately, one was ruled legal and the other was not.
usef- | 6 hours ago
From what I know, the "books3" dataset was normalised in the LLM and research ecosystem, where collected datasets were seen as valid to train on and/or fair use. I'm not sure any of the major frontier companies are free from that, if we don't believe it was fair use.
I do think most of their choices are explainable by "they just believe in agi risk". You truly wouldn't want non-agi-pilled companies to train on your data and approach the frontier if you were worried. You might slightly hurt your own business with safety filters (that no one else does) if you were worried. They are less worried about other "moral" decisions like "sharing" if they conflict with AGI: the research they still share is all of their safety research.
This definitely doesn't make them "good", but they do seem fairly "consistent". Most of these issues were talked about publicly by the founders long before Anthropic was founded and/or the AI race+money appeared.
8note | 6 hours ago
thats not something you expect from a company that "believes in agi risk"
usef- | 6 hours ago
Note that the companies that haven't faced these issues so far are the ones that don't do safety testing, or don't have frontier models. I'm not sure who I would pick as "better" on any of this right now.
janalsncm | 6 hours ago
That doesn’t mean I like them pirating books and being shady about tokens and paternalistic “safety”
marcuschong | 6 hours ago
badsectoracula | 7 hours ago
He wants to build a tech-god kept in chains whose power he parcels out to the unwashed masses he deems worthy like some sort of high priest of intelligence.
And that is being charitable and going by the interpretation that he actually believes what he says.
im3w1l | 7 hours ago
spiderfarmer | 7 hours ago
dimgl | 6 hours ago
jansport123 | 6 hours ago
greatgib | 6 hours ago
React for example.
And we could easily guess that there wouldn't have been so much open source models, and grand public experiments and free tools if llama models were not release to the general public.
inferniac | 6 hours ago
anukin | 6 hours ago
ballon_monkey | 6 hours ago
Edit: I get it. It's a hard pill to swallow. I understand people don't like Musk or Zuck. But it doesn't change the fact that you're being lied to and brainwashed.
souvlakee | 8 hours ago
jumploops | 8 hours ago
The model seems on par with Sol and Opus 5 on paper (admittedly on some older/saturated benchmarks, but very competitive for $).
Stats:
1M context, $0.10 input/$0.002 cached, $0.20 output (Mtok)
2001zhaozhao | 8 hours ago
(It's probably going to be a bunch of repetitive batch jobs like web search that have no training value)
dbbk | 8 hours ago
winstonp | 8 hours ago
hadlock | 7 hours ago
HDBaseT | 5 hours ago
Muse Spark 1.3 supports Text, Image, Video, File, Audio inputs. We've only started to see models from China include image and video inputs recently.
a012 | 3 hours ago
fibonacci112358 | 8 hours ago
IshKebab | 8 hours ago
r_lee | 3 hours ago
cnxhk | 8 hours ago
lostmsu | 7 hours ago
Bolwin | 5 hours ago
dangoljames | 7 hours ago
LZ_Khan | 7 hours ago
maciejgryka | 7 hours ago
apodolny | 7 hours ago
mmastrac | 7 hours ago
gehsty | 7 hours ago
Only way I see is if it becomes the new SOTA / frontier, does anyone think Meta will surpass Anthropic or OpenAI?
I still can’t get my head around why language models are an existential threat to Meta - they own the platforms people watch adds on?
phyrex | 6 hours ago
improgrammer007 | 7 hours ago
jmward01 | 6 hours ago
popularonion | an hour ago
We understand theoretically they're taking our data, but yeah, that data is vital to the entire business plan of all these companies and WAY more valuable than people are giving credit for.
I checked up on Mistral recently and saw their Claude-alike coding harness is using GLM now, whatever it takes to keep users on their platform and feeding them data.
anjel | 6 hours ago
ryanschaefer | 6 hours ago
Aurornis | 5 hours ago
coolcoder613 | 5 hours ago
dcl | 5 hours ago
alexboehm | 5 hours ago
dcl | 5 hours ago
This kind of raises another question to me regarding the coding benchmarks, how much of it is model versus harness?
wkcheng | 5 hours ago
dv35z | 4 hours ago
yanjunnf | 4 hours ago
keyle | 4 hours ago
m00dy | 3 hours ago
israrkhan | 3 hours ago
esafak | 2 hours ago
ydna404 | an hour ago
MitziMoto | an hour ago
IIIIIllIIII | an hour ago
geoffbp | an hour ago
This is interesting
water-drummer | 11 minutes ago
Iolaum | 10 minutes ago