My somewhat uninformed guess is that the weights of the target network (Qwen or Gemma), when they reach the context of the larger editor model (Astra), the attention process is doing some sort of PCA on the weights, conditional on the prompt ("make this better at games"), fetching the relevant weights from its own network (which serves as the larger database), transforming it into whatever latent space the target model is amenable to and pasting them. If this is the right intuition, it is amazing what attention is able to achieve. As a test, I think smaller networks won't be able to make big ones better using the same process.
EDIT: I think the bigger implication is that backprop is not the only game in town.
I think they're saying this is all happening in latent space through pure reasoning, and "fetching" is just a metaphor for that, not a literal tool call. I do believe it's possible for models to learn the capability to "introspect" their own weights during training.
Yes, I should have been more precise - I meant the information embedded in their weights (which every neural network retrieves to function), not their literal weights.
I don't know if some of the labs already knew of this. But I think there would be a limit to how big a model you can push into the context of the editor models. Typically, with 1M context and say 5120 hidden dimensions supported at fp16 / bf16, you could potentially edit the weights of an fp4 20B param model (roughly).
As many have mentioned, BG3, CDPR's Witcher and Cyberpunk 2077 franchise, Remedy's Control / Alan Wake franchise (both original IPs) to name just a few. DICE (within EA) comes from Sweden and it's responsible for EA's core Frostbite engine alongside its Battlefield franchise.
I cannot point to the reasons as to why. Many of these teams are global teams - so a lot of the same problems apply: marketing costs/timelines, release-time crunch etc. The one difference I can point to is that perhaps executive payouts are not overly dependent on total shareholder returns.
I am using it via OpenRouter for going through some Vulkan code. It's a beast - it has no hesitation putting in instrumentation within the code (having a number of render passes) to dump the framebuffer to PNG files and doing pixel-by-pixel comparisons, putting in false colors in shaders to identify bugs etc. Very impressive.
That’s something Opus can also do. Perhaps by coincidence I also recently debugged shader bugs where Opus dumped the resulting render to PNGs and then using Python PIL to do comparisons. It purposefully added false colors too.
Suspect there is a training set going around focused on comparing and extracting data in distinct steps from image input. Saw it first on Opus 5 on a backpack construction task to extract and compensate for a reference tape being slightly offset in a LIDAR scan, followed by heat maps for per pixel comparison between the reference and blender render. Opus 5.5 did go a step further in the same, extrapolating offsets to compensate for capture inconsistencies. Some may not even notice this happening as Claude Code steers the model to use the scratchpad hiding much of this work.
Found this especially impressive because backpack material related research has been something I know models can still get caught up on and the models being decent at web research in coding nowadays does absolutely not apply to these task equally Recommending nylon canvas shell suppliers or waxed cotton when PVC coated cotton is explicitly request, that sort of thing. While not perfect in reading coding docs, models are far better there.
Like with front end and code “parsability”, feel like you can very clearly point at which labs use what training set, with many often sharing the same suppliers. Once I noticed that, I became very bullish on specific training sets over emergence as the source for current gains.
Out of curiosity, which harness do you use for this kind of task, and how do you sandbox it? I used Aider before, and consider switching to a more recent tool.
I started out with agy CLI but it's horrible because of its very indeterminate context compaction. But I live with it because it's the only way to use Google's subscription. I was and still use Opencode in parallel (accessing Openrouter ZDR endpoints) when I run out of rate limits / weekly limits.
Over the past two weeks, I have started using a locally modified version of deepseek-harness (modifications include disabling telemetry and querying a provider's model list dynamically through Openrouter, instead of relying on the pre-packaged pi-ai manifest that comes with the base code). So far, that has been the best experience - I really like the caching, tokens/sec stats that it provides. (EDIT: I don't know a lick of TypeScript but AI-driven modification of the harness is good enough for me to make these tweaks).
Sandboxing is a definite issue for me. As you perhaps know already, headless operation is often not an option in graphics/game programming which requires actual render-surface instantiations from windowing systems and then input-context handling (keyboard/mouse/gamepad) as well. So I run most development on a VM. But part of my test-suite and smoke-tests rely on runninng it directly on the host OS (which is not ideal). I'd love to hear suggestions to avoid this.
> So I run most development on a VM. But part of my test-suite and smoke-tests rely on runninng it directly on the host OS (which is not ideal). I'd love to hear suggestions to avoid this.
I see Typeface AI have 30-40 people working for them on LinkedIn (some are VC advisors / board members), a lot of them being engineers. Their openings suggest a strong developer market focus (to begin with). I don't necessarily see a lot of people at the company with enterprise sales channel experience, but they already claim to have several Fortune 500 companies as clients (maybe the VCs are helping there or there is natural dev-driven traction).
So they clearly have a product, a strategy around it and perhaps the compliance scaffolding (SOC2 Type II etc) that may be needed before actually being able to charge money for it. They also have the right brand names associated with the founding team. Execution, so far, seems good enough to create a splash, at least.
As an investor, the question(s) to ask is (in my view): "How do they make money? Will that way to make money survive?". The answer to the first: selling input tokens and perhaps subscriptions/credits eventually.The answer to the second: "Yes, but with the risk of unit revenues declining faster than their unit costs". How can they mitigate this problem: by being big (scale / mindshare etc) so that their unit costs (including for customer acq) fall faster than their unit revenues will - I believe that is the question most AI companies are trying to tackle these days. Any new competitor will have to tackle basic fixed costs (of getting started) first before even getting to the stage of having the luxury of worrying about unit-economics.
So yes, they might eventually be competed away but whoever is in their team is trying hard to make a useful product/ecosystem and that should be applauded, not ridiculed with "it's all marketing". This is way more than a simple github/huggingface-based open-source replica solution can hope to achieve without institutional backing (either big-tech or system-integrators).
What should rightly be questioned, of course, are the valuations the VCs are providing to them in hopes of passing this hot potato to a willing buyer (say a hardware maker like NVidia) - the incentives there are very well defined and depend very much on perceived TAM (which lately is on very shaky ground given how far token pricing has fallen causing, among other things, OpenAI to "miss" on the market's expectations for annualized revenues, even before they're listed!) [1]
URL? I can try to take a look and pick 2 or 3 I like, as a recommendation to post here. (Anyway, it's difficult to predict what the hivemind will upvote.) (Which are your 2 or 3 favorites?)
It's astounding that the thought to dissociate one of the mathematicians from the proposed publication was driven by their corporate institutional affiliation - and that that exclusion was suggested by a scientist themselves! This is like a researcher from CMU saying to an NYU researcher that their collaborator, being from MIT, is a problem - this is as ridiculous as that!
Progress in humanity's knowledge now has to play second fiddle to narrow corporate interests as IPO timings near (both of which wouldn't exist anyway if generations of mathematicians hadn't paved the way for AIs to become as good as they have).
The scientist allegedly making that request comes from a machine learning background. Perhaps he's not familiar with the culture in mathematics regarding authorship. That sort of squabbling over author priority would be unconscionable to mathematicians.
Thank you. That indeed appears to be the best approach.
Judging by some other comments here, I sense many people here (just like much of the internet) are just angry by default for some reason and assume no prior effort on the part of the poster rather than offer solutions to what is a legitimate business problem.
There are two fundamental problems: one is the inability to admit mistakes and shed the false aura of omniscience in front of customers, investors or employees. The second (almost a corollary of the first) is overpromising and, as a result, underdelivering because those who face the customers/investors want to appear infallible.
EDIT: I think the bigger implication is that backprop is not the only game in town.
reply