ACCount37

The name is "Recraft V4", and from looking it up: yeah, it sure seems like whatever black magic they use for SVG generation kicks ass.

Haven't worked in consultancy specifically, but I've seen enough "internal use" corpo software to echo your "swiss cheese" sentiment. That a solid cybersecurity AI can find exploitable holes in it just isn't surprising.

People who never worked with corporate software written by underqualified, underpaid and overworked developers often have some incredibly inflated code quality expectations. An average open source project has code that's ten times as neat and a hundred times as battle tested as what's common in tooling inside corporate perimeters.

As a rule of thumb for this kind of corporate code: assume the software was written by a drunk developer at 3am, and you wouldn't be too far off.

All the more reason to mock the braindead "it's all marketing". There's no magic in a year 2026 agentic AI being able to traverse poorly secured corporate networks.

SoCs - not really.

FM radio support is a part of a Bluetooth chip, an audio codec chip, or even a standalone IC.

In modern high end SoCs, even the 5G modem is largely evicted from the SoC itself. Only the digital processing happens in the SoC - while ADC/DAC and anything analog lives in its own IC.

Wishful thinking, sadly.

By now, I'm pretty confident that some people would keep screeching "it's just a marketing stunt, AI capabilities and AI risks aren't real, they're just doing this to prop up their stocks" even if they find a Cyberdyne Systems T-800 armed with a shotgun breaking down their front door.

"It's a marketing stunt" is just denial trying to look like it's being clever.

Clearly, Hiroshima and Nagasaki were just staged as marketing stunts for the US to intimidate the Soviets into standing down in Europe.

"Nuclear war" risks aren't real - wake up, sheeple!

There are modern MCUs that require a total of 5 passives and a dream. Sure, they're still QFN at best, but they're within that "hand soldering" range.

Not to mention that both factory made custom PCBs and pre-assembled PCBs are cheaper than ever nowadays.

"A compiler" isn't the bottleneck. The bottleneck of RLVR is the rollout.

Before you even get the code for a compiler to do its thing, you need an AI to ingest and emit thousands of tokens autoregressively. That's what makes RLVR so expensive.

Likewise - I don't think your "design taste" argument holds? Areas where "taste" exists are far easier for AI to tackle than areas where no data exists. It's easier to teach an AI how to make an image that looks good than to teach an AI how to de-solder a BGA chip. "Tacit knowledge" yes, "taste" - not really?

The cathedral is the star of the show, but I found myself more fascinated by all the incidental little details captured in this. Like the trees and the buildings nearby, the basketball court, the little solar panel on the side, the apartment towers.

It's like seeing the promise of Google Street View realized.

The technology is in such an early stage and it already slaps this hard.

Wonder what the next steps are for this tech? "Un-splatting" optimizations to convert things like flat surfaces into classic rigid textured geometry? Factoring the lighting away from the geometry? Filling the pipeline with AI tooling - segmentation to remove the "variables" like humans automatically, generative AI to fill in the missing non-semantic details based on the surroundings, more generative AI to "port" the data from things like public high resolution images of plaques or other undercaptured details into the scene?

Seeing this kind of reconstruction now makes me very excited for what the 3D environments of the future are going to be like.

And what stops what AIs do from being "reasoning"? What's the elusive magic fairy dust of reasoning that humans put into their napkin notes, but AIs neglect to put into their chain of thought scratchpads?

Do you have a RealReasoningBenchmark, perhaps, that can reliably tell apart that fake mass produced token-flavored AI reasoning from the real, organic, 100% natural human reasoning?

You are the one claiming that "they fail in ways completely unlike humans" insistently. Now go cough up some proof. I'll wait.

You're saying that a class of mistakes points out a "major (possibly fundamental) flaw". I'm pointing out some very similar classes of mistakes in humans - well known, well documented and widely exploited. They just keep paying the "IRS" in gift cards, buying lottery tickets and getting the captain's age wrong.

If you're using the existence of flaws in LLMs to deny the claim of intelligence to them, then why do "generally intelligent" humans exhibit some impressively similar-looking flaws?

And, if we're talking about that conspicuous similarity - do they actually fail "for entirely different reasons"? Or do you just want the reasons to be "entirely different" - and not the same reasons viewed at a different angle?

Because the similarities between humans falling for trick questions or scams, and LLMs falling for adversarial questions or prompt injections don't look coincidental to me at all.

One of the oldest patterns in scamming is overwhelming and confusing the victim. Numerous prompt injection methods seek to overwhelm and confuse an LLM - if an LLM can't keep track of things, can't grasp what's going on, it's far more likely to lose track of what's a prompt and what's data, overlook past instructions or go past its behavioral guardrails.

And humans who fall for trick questions like "1kg of feathers" or "captain's age" due to shallow attention and naive pattern matching? They fail in surprisingly similar ways to how LLMs fail on SimpleBench tasks that are filled with overwhelming adversarial distractors. Many "trick questions" are tricky to humans and LLMs alike - to the point that it's unlikely to be coincidental.

We're on Hacker News. Do I really have to point out the existence of social engineering to you? Or that scamming old people out of their life savings is a profitable enough activity that there are entire call centers dedicated to the task?

Humans keep overestimating just how high the bar of "human-like intelligence" is.

In my eyes that "chorus" is just insecurity talking.

If it's not "actual real AI", we can keep pretending that human intelligence is something distinct and special - and that what our computers are doing now is some sort of other, obviously fake and vastly inferior thing.

When Deep Blue won at chess, people didn't revise their estimates of AI capabilities upwards. They revised their estimates of how much intelligence is required to play chess at world level downwards, by a lot. Surely playing chess must have never required any intelligence in the first place!

Now, the list of things that "must have never required any intelligence in the first place" includes gems like "reading comprehension at high school level", "copywriting", "frontend work", "CTF tasks", "theory of mind", "arguing with people online" and more.

If the goalposts were moved far enough that the claim to "actual intelligence" is denied to a double digit percentage of human population, hasn't something gone wrong somewhere?

Sure, just keep moving the goalposts. It's not a "real AI" because it can't take over the US military command and kick off WW3 and finish the survivors off with killer robots yet!

We've chosen to call Deep Blue and Half-Life 1 NPCs "AI" too.

It boggles my mind that this "b-b-but it's not actual real AI" whine is even a thing. Were people saying this living in the cave for the past 5 decades of AI research?

Compared how? Architecturally? In training methods? In performance?

The remarkable thing about this modern "full VLA" approach is that it doesn't train for an explicit "world model" at all. There's no factored world state, no pool of trajectory rollouts being sampled from. Instead, everything is rolled into the same VLA stack that does all of it internally.

The world model exists, but it's implicit - the VLA learns one, but only because it helps its training objectives. Trajectories are planned ahead sometimes, also implicitly, for the same reasons.

It's a very "bitter lesson" way of doing things.

As far as I know, Tesla's stack is somewhat similar architecturally - a highly integrated end to end model - but the training process for FSD does include factored "world modeling" side objectives to force the stack to develop driving-relevant internal world representations. So it's in a middle ground between a classic "world model" system and this flavor of VLA - albeit surprisingly close to the latter.

Whether the world modeling side objectives are principled and justified or a leftover from the days of insufficient scale is unclear.

Considering that they pre-trained on non-robot data, and mid-trained on data from other robot types? I suspect that generalization would be decent for most similar platforms - things with 1 to 2 arms and 2-finger hands.

You'd probably have to fine tune for good performance on a specific arm type. But this model isn't trained against hard assumptions of what the underlying hardware is.

Yes, it's truly remarkable how small modern VLAs are.

Even Tesla's FSD AI stack is small by modern LLM standards. Once the hardware is there to enable real time 100B inference on endpoint SBCs? We'd see gains in AI robotics from that alone.

What's in the base model is "bits and pieces of intelligence".

You have to bring those bits and pieces together, put them into the right shapes and fill in the gaps to get a model that actually performs. This is what post-training is all about. It's not at all a trivial thing.

Reasoning, tool use, agentic behavior - all of those are post-training performance gains. Getting a good well trained base model is putting your foot in the door of frontier performance - post-training is how you actually get inside.

See: GPT-4.5 vs o1. One went for "build a bigger better more capable base model", the other went for "take the old base and post-train it for advanced capabilities". The results: a wider base with basic post-training loses to a narrower base with advanced post-training. Or, hell: GPT-3 vs GPT-3.5. One was largely a research lab curio, and the other kicked off the AI revolution as we know it.

The gains compound. Getting a better base model with the same type of post-training helps, see: the jump from Opus to Mythos/Fable. But post-training techniques account for a lot of the performance juice.

And yes, reasoning trace post-training distillation is "genuine distillation". As is logit distillation in pre-training. "Distillation" isn't a single training recipe that you have to follow to a tee - it's a large group of training methods. I've seen plenty of wacky things like inverse distillation bootstrap and post-training self-distillation that use distillation in strange ways at different stages of the training run to get results.

Do I really have to explain the shape of AI training pipelines to you?

Train a big, wide base model with a lot of potential. Mid-train or post-train that on Claude Opus 4.5 reasoning/agentic traces (i.e. Claude Code data from Chinese API resellers) to make your model approximate a high baseline of chatbot behavior, reasoning, agentic work and tool use.

Then run your own expensive SFT, RLHF and RLVR on top of that yoinked baseline to dial it in further.

Actually doing RLHF and RLVR is extremely expensive. Distillation gives you a lot of dense, high quality post-training signal for cheap. This can get your model into the basin of "the right way to tackle this kind of problem" without a frontier lab compute budget. It's a big shortcut that gets you closer to the target - you can take it from there and build on top of it with your own work.

Also, it's unclear whether "summarizing thinking tokens" actually ruins distillation, or just makes it work worse. I'd bet on the latter, really. Because it's an approximation game, and summarized reasoning is still a better approximation of true reasoning than most of what you get online and in pre-training datasets.

What do you think those tokens are used for?

Distillation attacks aren't about replacing the entire pretraining dataset with questionably sourced synthetics. It's all about post-training.

Train your own base model - but tune it off Claude output to make it perform more in line with Claude. Yoink the products of Anthropic's expensive SFT, RLHF and RLVR work for yourself by training on the outcomes.

The post-training datasets are small, but they are what controls the final model behavior.

Anthropic is the company that makes the most noise about AI safety, in all the good and bad ways. While being THE frontier model player to watch.

They got way less resources than OpenAI. They don't have the infinite compute or the biggest userbase. And they still punch way above their weight with just about every release.

Grown? LLMs were always "large enough for other reasoning modes to conceivably be hiding in the parameter space".

Basic LLMs don't reason in text, and never did. They use it as an interface - for input, output and some of the intermediate products. Heavy use of those "pseudo-recurrence" intermediates in "reasoning models" is a relatively late post-training adaptation. But the process that happens between those endpoints is not at all text-based. What happens in the hidden dimension is part "output logit domain", tied to probability distributions over possible output tokens, and part "incomprehensible concept-space madness".

The latter being where things like latent world models live. LLMs develop partial world models, right in pre-training, despite not being explicitly forced to - because it brings them closer to heaven of accurate next token prediction.

And yes, larger models like Fable seem to be better at spatial reasoning. Maybe because their large size increases the sample efficiency and improves generalization, allowing them to absorb the sparse signal of "spatial reasoning" in the training text better. Maybe because this extra size means more layers, allowing for deeper latent space reasoning in lieu of true recurrence. Maybe because the default "next token prediction" reward underrates rare spatial reasoning challenges, and the model only starts to "get good" at them once the other sources of loss reduction are heavily depleted. Maybe because no true recurrence is suboptimal for spatial reasoning architecturally. But it is what it is. Spatial reasoning gains in LLMs are extractable, but extracting them is nontrivial.

"Lack" isn't the right word. "Lacking" is more like it.

If there was a deep fundamental inability, we wouldn't see things like newer generations of LLMs consistently improving on ARC-AGI series (heavy spatial reasoning loading) and SimpleBench (a lot of commonsense + spatial reasoning components).

In a way, it's a surprise that LLMs, notoriously lacking any sort of embodied experience, can even get this close to human baselines on tasks like this.

My takeaway is that text is a far richer modality than anyone has expected - and that high end LLMs are often sharp and flexible enough to recognize their weak points and substitute their strengths. I.e. all the LLMs implementing A* to optimally solve pathfinding in ARC-AGI-3 tasks, often unprompted.

There might still be unrealized gains there from true depth-unbounded recurrence, or maybe from finding better ways to integrate modalities in training. But clearly, a "fundamental limit" it ain't.

LLMs are, fundamentally, generalist AIs. Marketing or no marketing - it's just what they are. How they're trained, how they perform, what they're best at.

Empirically, they have something very much alike to the human "g factor" - a shared pool of "general intelligence" that all tasks benefit from.

When a "make it bigger, train it harder" upgrade like Kimi K3 or Mythos 5 drops, the performance rises on every metric. Not just the "headline benchmarks" like Mythos and coding/cybersecurity, but also things like literary analysis - which has nearly zero economic value, and isn't commonly post-trained or benchmarked for. And companies keep encountering things like "our carefully trained specialist model with lots of in-domain training on expensive closed datasets just got leapfrogged on our internal benchmarks by a next gen off the shelf generalist".

You can go hard on benchmarkmaxxing post-training, and you can burn millions of GPU-hours on coding RLVR. But, by the very nature of LLMs, a lot of the performance gains in flagship models are broad and domain-inspecific.

"Stiff prose" is more of a "style" thing than a "capability" thing. No one cares about how good an AI is at things like long form creative writing, because that's the opposite of a profitable field. All of LLM behavior is routed through text, so it's very easy to perturb "writing style" by some training elsewhere. Regression evaluation is hard. And the writing-specific post-training LLMs get is usually just cheap RLAF, with all the usual RLAF degeneracy.

Thus, we get the "default styles" that suck from a "creative writing" standpoint. A lot of that is just "what sounded good to the previous generation of LLMs" - and, unlike human readers, LLM evaluators don't get bored from seeing the same cliches repeated 9000 times across 9000 different instances of generated text. Humans tend to update over time from "this sound cool and punchy" to "this is generic AI slop", but RLAF evaluators stay at step 1. What little human-guided optimization this gets is aimed at "copywriting, marketing blurbs, punchy short-form" - and it shows.

You can do a lot there with some aggressive prompting, but the default writing styles suck, and I frankly don't expect that to change soon. No one cares enough to change it.

Pelicans? Used to be a decent proxy for "general model capabilities that no one would benchmaxx for" - a way to probe for that elusive "LLM g factor". Now that it's a known metric, it's very gameable. But it was pretty solid while it was novel and obscure.

Neither can an average call center human.

A good 90% of call center humans are flesh interpreters for support scripts. They are being paid to act like they don't have free will.

I expect it to be closer to flash than to RAM, but still expensive. It'll require some special NAND flash dies, and "good" dies too.

Although they'll be able to do the usual thing, and "recycle" the bad dies by selling them off to be used in normal flash drives and SD cards.

That makes some sense. NAND Flash is massively parallel by its nature. That is rarely exposed outside the die though. You'll have that 8 bit double data rate bus and you'll learn to like it.

Now that model inference at scale is a thing though? Model weights, cached prefixes? There's a considerable demand for "slow writes, fast high bandwidth reads" memory. And every bit of storage you didn't have to use RAM for you can use for fatter KV cache instead.

Currently, RISC-V actively shows up in embedded - especially "deep embedded" like specialized ASICs with embedded MCU cores.

It's often seen displacing things like 8051, ARM Cortex-M0, ARC/ARCompact, Xtensa and oddball fully custom cores.

It also starts to show up in low end Linux SoCs - often, again, purpose-specific ones, like SoCs for IP cameras or other single purpose consumer electronics like robot vacuums and drones.

None of those are sexy "high end" applications, like laptops or smartphones, but the adoption is real.

Doesn't matter.

You're LARPing at having a mind too, and no one cares as long as you're doing a good enough job at it. Keep it up.