2001zhaozhao

There's now a 40X discount in the cache input pricing instead of 10X.

This seems to point to them having achieved some kind of optimization in attention mechanism perhaps along the lines of DeepSeek V4, which had a similarly high discount between cache input and normal input.

In real world use, the savings should be quite noticeable. For example, you can now use the model at 800K tokens context window at the same cost efficiency as the previous model at 200K tokens context window.

Annnnd this is why we can't have nice things

It's really about storing institutional context and on-task learnings. The AGENTS.md can do the same thing as memory, but if you have a good memory system, in theory you never need to do any ongoing maintenance of AGENTS.md and the system will do it automatically. And because it's automatic and has a whole harness system built around maintaining the memories, you can engineer it to have not just more convenience of maintenance but also better correctness / recall quality.

Memory systems can also let AI load memories on demand so it's actually analogous to maintaining AGENTS.md plus twenty different "read this if you need to do X" markdown files.

I think the name "memory" makes sense given that the AI system is one singular system with central context (rather than a software org of multiple distinct humans with distinct memories), so the equivalent of institutional knowledge in the software org really is just akin to memory for the AI system.

If the counterargument to knowledge graph-based memory systems is that they're slow and take multiple steps, then it's not really a counterargument. I'd happily trade off speed for giving the agent ability to find more precise memories.

That said, I think this article's basic idea of "text files + semantic search index" is a good way to implement memory because AI is already good at search, and semantic search is far more flexible than a knowledge graph and decreases the chance of the agent simply not being able to find a memory (or inserting duplicate memories, etc.) due to deciding to go into a slightly different branch than the one it needed to travel.

That said, I'd probably layer additional steps on top for further efficiency improvements. For example, automatic concise summaries of memory items, other supplemental indexing methods for better memory recall, etc. My ideal solution would be multi-step and therefore wouldn't solve the knowledge-graph's slowness, but it would solve the rigidness problem of a knowledge graph.

I would love to see models that can think at different rates and also output a thinking scratchpad alongside output text instead of before all output.

Right now models need to rely on less legible compressed CoT to get high intelligence per token/step, but with diffusion they would just need to output more tokens per step instead.

> The biggest shift for workers will happen when AI provides nearly error-free work. At that point, it will be able to function on its own without a human checking in on it, and companies will have every economic incentive to let it.

This is a very succint argument that has a technical implication - that correcting errors and misalignment is incredibly important for autonomous AI capabilities in a fundamental way.

A dream of mine is to be able to host a LLM-powered video game that I can host on a home server running a decent mid-range GPU like the RTX 5060, and the LLM is fast and intelligent enough to make for a fun game experience for a few dozen concurrent players. People can ask for features and they just get made and added to the game on the fly for the lobby to enjoy. The hosting costs would be manageable enough that I don't have to charge anything for the game.

I think with one more year or so of small model progress, that might just be possible to accomplish.

I've played around with Mindcraft for a while. Fair warning that it is quite outdated and spaghetti-coded. Although all of its components required to make it work are somewhat hard to replicate from scratch, so it might still be your best option.

I like that to type s****** you had to type s************.

Gemma 31B dense at 3k TPS seems like a pretty big deal. Groq has always been more efficient than Cerebras at running very small models and this release seems to be no exception.

ooo, nice Workflows feature you got there.

yoink

It's interesting you brought up Lore. I also wonder whether importing VCS tech from the games industry for versioning general agentic work would be a good idea.

I was just gonna blast them for the infinite promo extensions and Fable, but at least they apparently have noticed this indecisiveness problem themselves and making it permanent is definitely the best way to solve it, lol!

After they perma-extended Fable into including it into the max plan, I'm willing to give them the benefit of the doubt with this one.

I think we need a similar technology to World being built by a more trustworthy company. The question is how trustworthy does one need to be for people to trust you with their very online identities that are literally needed for others to recognize them as human?

I suspect that this kind of tactic is going to be everywhere in a year or so. Entire fake personalities and organization websites on the Internet created just to push a narrative or to advertise a product, which completely drown out real information.

At some point they may be indistinguishable from human-produced work, so the only way to verify if a site is someone's genuine opinion or AI-generated narrative is through some form of authority (or something else that is very difficult for AI to fake, but that protection can be broken by advancing technology).

If that happens, then in turn AI chatbot makers would in turn need to look for these kinds of authority to continue to provide accurate information, and said authority websites can use their privileged position to charge for access of their data.

On the other hand, it used about the same tokens as GLM 5.2 and got 1 point lower score.

The fact that we have a GLM 5.2-class model that can run on two 3090's comfortably at Q8 is absolutely insane. It wasn't long ago that GLM 5.2 was considered amazing for open weight models.

bruh

I sometimes run into browser JavaScript GC issues for browser games specifically but on the JVM side i have not run into any pain points for a long time.

(I run a first-person shooter Minecraft server, and for this and other fast-paced gaming in general pause times under a millisecond or so is generally good.)

They're nearly certainly going to use it internally to speed up research that is serially bottlenecked. I would bet this is why they're interested in the Cerebras partnership more than everything else

Early in 2026, Vibe Kanban tried to do a multiplayer kanban board to code on a project collaboratively where each individual task is still managed by one developer and their agent.

I find that to be a much more promising route than Delta's straight-up multiplayer development on the same task. (Vibe-kanban did fail but I don't think their idea was wrong, but rather the execution- they didn't differentiate enough from straightup Github and Jira boards when that was easily possible, just centralize the agents along with the board.)

Not that two people being able to look at the same agent at the same time is bad though, it's probably amazing for pair programming especially in a remote work setting. It just shouldn't be the main point of a product.

I also think that Delta's codebase replication approach is just flawed in general. Why would you build a remote server just to sync the codebase between multiple developer's laptops when you can just put said codebase on said server in the first place? No expensive sync is needed and all you need is to do remote multiplayer access at the UI level, and everyone gets to work happily on the same codebase and talk to the same agent at the same time.

To me what matters is being able to verify human authorship at a glance. It used to be that when you saw a work of visual art you were automatically 100% sure that someone had to have spent a great deal of effort producing that, but image models took it away. This is the same situation but with game environments.

This kind of content generation can be quite cool to enable new game ideas for indie developers that were only possible with AAA before. But it also makes it hard to gauge the amount of human work that went into the game. When you see an elegantly designed building, environment or easter egg in the game, was that really made by a human (or even prompted by a human), or did the AI generate it autonomously as part of a much larger generation step?

I think this is overblown, the watermarking is just enforcing a particular recognizable output text style.

The existing Claudisms like "smoking gun" and "it's not x it's y" evidently haven't impacted the model's agentic capabilities much, and i doubt the new ones will either.

Is saying "XXX: An Agentic Development Environment" already sufficient to get to the front page of HN? I guess no differentiation is even required in this category, huh?

In the case of my WIP Agent IDE, OSS is mandatory because the tool needs to be able to read its own source to improve itself.

i'm looking forward to Qwen3.8 27B launch to see how much models have peaked at a given size.

it might already be time to start burning the best small models onto hardware since it's possible they can't get much better at many tasks like knowledge recall due to the inherent information density limits for models at a given size.

There are always uses for outdated models.

Claude Code is still using haiku 4.5 from ages ago for explore subagents for instance. Not to mention production uses like customer service that only need to be "good enough"

The question is what they plan on doing with all their scientific knowledge once they succeed.

Sure, they'll keep it internal for a while to make sure their knowledge bank is more thorough than everyone else's, and because oftentimes discoveries can be far more convincing internally than externally (you need fewer sigmas for it to update your belief in a certain direction). But then how do they intend to profit from it in the end?

> Note that Jeff and crew have cleverly structured their company to avoid problematic uses of AI

Do you have more sources/info on this?

I'd intuit that for an AI video editor workflow you'd probably want something in the middle of this CLI solution and a full GUI? I'd take a crack at making one if only i had some expertise in the field.