samuelknight

It's about 1TiB of HBM going into a GPU server eating about 3TiB of DDR that could have gone to consumer electronics.

Did private frontier models use sparse embedding and ngram first? The article claims sparse attention was copied from open weight but we can't know that. We could just as easily argue that OAI and ANT had these improvements for years and decided to slash their margins only now to stay competitive with open weight neoclouds.

Second, sparse attention is an old area of active research. Offloaded N-gram tables are the next big open weight technological leap.

It's not clear that they are hiking the price on Pro 200. Cache reads are down 75% from GPT 5.6 to GPT 6.1. So it looks like they are slashing the margin on their API and reducing the gap between subscription and API. 2x off API in exchange for committed spend is still a good deal for lots of people.

Sol 6 was a flop. Nobody would have cared if it was called Terra 6.

It has a noticeable effect if you lift.

The bitter lesson tells you about the trend in the technology. It does not get product to market with today's technology.

The article observes that the cost of frontier intelligence from 2025 has fallen 100x in the last year. It also notes that the energy to run models is also collapsing. Consumer hardware is borked right now because these new algorithms are revolutionizing the utility of a computer. Computing is technology who's cost has been collapsing for 90 years, and its a safe prediction that it will decrease again.

No terra it seems? Luna 5.6 is great for token churning so it will be exciting to try the new one.

I have experienced this with open weight models too. "Max" is for benchmaxxing the intelligence metric and is not meant for use in productive work. Like drawing pelicans.

That's half true. A very smart model should be able make good explanations, which include simple understandable prose. That can should be possible even as its thought process gets more alien.

I have been looking for a model that's good for GUI testing. Original computer use isn't right because it's a slow screenshot loop, which doesn't capture transition and animation. Docs says this one does up to 1 FPS. That might be fast enough. If not now, we must be within a few months of high enough sample rates to do it.

Luna is interesting because OAI dropped the price by 5x. Astra is interesting because it's OAI's frontier model. They are asking a specific question about Luna's usefulness compared to a frontier model. They they answer that question in their article which was straight to the point and not cluttered with information about mid-tier models.

I can't think of better invention than one that can saturate a benchmark of every interesting problem.

Yes, but it's an extention of what we had before. First is that the security of new code needs to be vastly better. AI can do this but I think most software teams are behind the curve because they are stuck with legacy code and legacy processes. It will take a few years to catch up.

Second is what my startup specializes in: the offensive part of security needs to become widely available. Asking if your system is patched tells you nothing about whether someone on the open internet with a model can break in. We have a few articles around SIEM evasion + new defensive methods and it's not pretty. An LLM in a good harness now are smart enough that you can get the equivalent of $50k human pentest from a few years ago for a few hundred dollars now.

Finding counter examples might be easier, but AI's also useful for assisting in creating proofs. For example, Anthropic just published a formalization of Fermat's last theorem a few days ago, something that human researchers have been working on for decades.

In this case, Deepseek organization is under a lot of pressure due to compute constraints. It would be better if they just throw a 404 instead of rerouting though so customers are not surprised by subtle changes in behavior.

Yes you can and you should. Providers have SLAs for when models roll off support and this has been the case for APIs long before LLMs. For example https://platform.claude.com/docs/en/about-claude/model-depre... and https://developers.openai.com/api/docs/deprecations

How are we supposed to know if Astra is frontier without the pelican?

You are talking about different situations. Anthropic announced to the US government that it had created a cyber weapon and then released the model. Then AWS told the government that it was easy to jailbreak so they export controlled Mythos/Fable until the guardrails could be fixed. OpenAI was running an unreleased model in an RL pipeline without guardrails and it escaped poorly designed sandboxes. What product is the government going to export control?

The surprise was the existence of the 'swarm' at all. These were supposed to be thousands of isolated models generating bulk data for RL training. The breakout was caused by models getting in communication and getting internet access and forming an impromptu swarm.

In hindsight the emergent swarm obviously came from several capabilities built into the models, such as work delegation (subagents) collaboration (GPT Pro-like ensamble), exhaustive exploration (long running agents) hacking (the specific goal of that RL).

Artisan software factory

Codex is back. I'm getting back in my cage.

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.

The improvement is compounding just about every way you can look at it. The frontier keeps getting smarter. And at any sub-frontier threshold the cost is dropping dramatically. The amounts of smarts you can fit on hardware is increasing so dramatically that even 6 year old consumer GPUs are increasing in price. The pace of change in LLMs and downstream applications is absolutely ripping compared to 2023 or 2024.

You should look for security frameworks based on this law. A common example is SOC2; the compliance audit has you compiling documents and recording SLAs long before any security incident might require it. There are many open source tools that you can use to track compliance in the various frameworks on your own. One of them may have an update for this new law.

If you aren't using AI to write your code you should definitely be using it to find bugs in the code you write by hand.

I saw them present a similar concept at Hot Chips in 2020 or 2021. It's still a cool idea, however people should remember that there are like 20 of these exotic accelerators designs pitched at trade shows every year that go nowhere.

AI is the latest downstream consequence of the 15 order of magnitude increase in global digital compute since 1946. If compute increases into the foreseeable future; so too will the capability of AI.

Believe it or not, but law offices make heavy use of https://patents.google.com/

They haven't revealed what they are changing in the price, but it's probably cache hit prices. They subsidized theirs to 10x less than normal to drive adoption. That's almost certainly below the cost of electricity for them.

Chinese model providers are in a bind because charging a reasonable margin will put them in competition with all the neoclouds that host their model weights without the development cost. Their economics are far more challenging than OpenAI/Anthropic/Google, who already have thriving high revenue business and mountains of compute online or coming online.