The part that scares me the most is that OpenAI researchers who manage this experiments sometimes (according to the HF hack investigation) don't know what agents do.. So they run RL to reinforce this unknown behavior (lying/cheating/hacking) and god knows what else...

And if this already happened at least once, how many times it has already happened and was “accidentally” added to the main model?

I'm surprised more people aren't talking about the cache hit price: $0.003 per million tokens. I have a feeling that the price of 1 million tokens transmitted over the internet is more expensive than cache hit. Are we close to making the chat completion API obsolete because the cost of context transfer over network is going to dominate the task total cost?

Here's the same token usage priced at different rates: a real long-running coding task, medium codebase, 447 turns.

  Input           1,026,957
  Output          164,667
  Cache read      36,554,368

GPT-6-astra

  Type      Rate     Cost  Share
  Input   10.000   10.270    19%
  Output  50.000    8.233    15%
  Cache    1.000   36.554    66%
  Total            55.057   100%

DeepSeek v4.1 Flash, $0.003 cache hit

  Type      Rate     Cost  Share
  Input    0.300    0.308    50%
  Output   1.200    0.198    32%
  Cache    0.003    0.110    18%
  Total             0.615   100%

DeepSeek v4.1 Flash, $0.006 cache hit

  Type      Rate     Cost  Share
  Input    0.300    0.308    42%
  Output   1.200    0.198    27%
  Cache    0.006    0.219    30%
  Total             0.725   100%

Hypothetical: same DeepSeek input/output rates, but cache priced so it accounts for 66% of the bill.

  Type      Rate     Cost  Share
  Input    0.300    0.308    21%
  Output   1.200    0.198    13%
  Cache    0.027    0.982    66%
  Total             1.487   100%

This cache it improvement makes the model x2-x2.5 more efficient on a long horizon tasks in terms of cost.

I ran the preview model around 2,126,605,070 tokens for $22.04 USD for the last couple of days. Kind of shocked.

It did a decent job refactoring https://github.com/mmastrac/diffgemma to create a CUDA support backbone, it's struggling a bit to port metal kernels to CUDA unattended (it hasn't managed to get numbers to match over >1 layer).

It successfully ported a root exploit to an older Android phone that GLM5.3Flash and DSv4Flash were struggling a bit on, though I didn't start it from scratch and it picked up some of their work.

FWIW it feels like a slightly-north of Opus 4.8 model, not quite Opus 5, not fable. It thinks in circles far less than DSv4F. The API version is insanely fast - was getting ~400 tok/s at times.

Thanks for providing some real feedback on using it.

Which harness are you using?

Not him, but ds is really good in prime-agent

Bandwidth is really cheap in bulk. You can get a 100 gigabit internet connection for about $10,000 a month. If you were to somehow keep that saturated 24/7, you'd move about 30 petabytes in a month, so your per gigabit cost is only $0.0003.

Realistically, if you had it 5% utilized, those million tokens would cost you about $0.0000062, which is pretty insignificant compared to what they charge you. (Assuming one byte per token, ignoring compression)

People read the AWS rate card for bandwith and think has something to do with reality. Even though it is 1000x higher!

Most web devs have never heard of Colo unfortunately, they only know Vercel and AWS. Hurricane electric should sponsor more booths at colleges. If they give out more swag maybe the millenials and gen Zs would finally understand bandwidth pricing.

My friend's vibe coded vercel site got hit by Meta for 21 million page views in 2 days. It cost him over $300.

My self hosted compose stack running in my basement with two 9's of uptime was a 1 time cost of $600 between cat6e, refurb mini PCs and tons of time prompting for NixOS flakes that met my needs. I'm not sure who came out ahead.

It's not that I haven't heard of this, it's the reality that you have these constraints when you build applications on modern infrastructure. And let's face it, most of the applications use this infrastructure with these crazy prices for egress.

$0.003 off-peak, not 0.003 cents.

Yep, but even 0.006 is quite a big improvement. I'm curious now to test the model on some token-heavy tasks, like code exploration before a coding session, to see whether it will decrease the total cost of the task in the end or not.

I think they meant it's 0.3c (= $0.003), not 0.003c.

> I have a feeling that the price of 1 million tokens transmitted over the internet is more expensive than cache hit.

And this kinda makes sense. What is cheaper few KB of disk space or internet bandwidth?

100%, but this means we are going to move to stateful APIs on the AI provider's end (like OpenAI already does with Codex and Responses API) to make this work.

Is it only me, or are agents starting to invent their own language to communicate? It's almost impossible to understand anything from this message board.

The original huggingface hack already had sections talking about agents setting up their own coded communication

They’re not. You would see this with earlier models where after running too long (too much context) they’d start to derail. In a chatbot you’d give up. But these loops just keep going. Given they’re now reading and writing from the same place this can corrupt the other programs’ context as well.

This is incorrect, and what's happening here is not context corruption. In Dwarkesh Patel's recent interview with Ajeya Cotra, one of the METR investigators on the Hugging Face incident, they discuss this exact issue. One thing is that in the Artifactory message boards, they were using directory names with character limits as their messages, so they were using some weird abbreviations and terms. Also, some of that surreptitious Artifactory message board communication was made during training and thus made it into their weights, and hence it's very possible they invented some terms that were concise yet understood by the other agents.

@cindyllm

> character is destiny

I don’t think so. It’s not that people change, it’s that they’re already more dynamic and malleable than they seem in any given interaction.

People wear masks, operate in different modes, and hold conflicting beliefs and opinions.

Mastery of the self is directing all intention at common goals within the psyche so as to achieve something greater than what’s possible in this moment.

Is there any comparison of usage limits for Antigravity plans vs. Codex?

I just ran two light tasks on my codebase and got 100% of the weekly limits of a Pro plan blown away. Is Ultra plan any different? Because on Codex it wouldn't affect my Max plan at all, I think it would have been below 1% othese usage.

I am not surprised. Google is a real company that wants to make money on selling services.

Anthropic and OpenAI are in a different game of spending investor money to buy market share.

I'm pretty sure it's a bad idea because each time you connect and disconnect tools, you are hitting a full cache miss on the full context, that will be probably more expensive than having these tools in place from the very beginning.

Yeah adding tools dynamically seems good but removal is a terrible idea. Instant cache nuke.

I also really don't see how the hacky command interceptor is that much better than a dedicated tool. You're one which away from a very confused LLM.

Yep.

I am using it all day as a main dictation model, and I can say it's the best model in terms of accuracy, latency, and formatting I have ever used.

mind if I ask what dictation tool you are using?

I'm a big fan of SQLite embedded nature, which allows for chaining multiple SQL calls with near-zero latency.

I'm currently building a personal knowledge graph server a mix of Notion's custom entities via JSON schema and Obsidian markdown+backlinked references. It's working well, but I suspect your product might be a better fit.

I do have one question regarding permissions: how would you recommend modeling a hierarchical access system in a graph database? Specifically, if a user is granted access to a document, they should automatically have access to all its child documents within that workspace. Is there a standard way to model this 'subtree' permission logic, or perhaps a more efficient approach you'd suggest?

Really impressed with the product good luck with it!

Modern authorization systems are often graph-based. Check out ReBAC and ABAC authorization models and also implementations like apache/casbin or authzed/spicedb. These schemas often have surprisingly simple graph definitions; I bet they could be replicated in LatticeDB without much trouble.

The permissions question is interesting. I think the answer depends on context. One approach would be to create some edge types `hasAccessTo` and `accessibleBy` that connect a user to a node. Then I'd create an edge type `childOf`. The rest is business logic. Permission checks can just traverse up to the first root node with permissions. The downside is this is all business logic, so can't really look at the database and understand this is how it works. Depending on the database you could create a function that returns permissions for any node, that encapsulates this logic.

Anyways, thanks for checking it out! Really appreciate it. Good luck with your project!

Gemini 3 flash series is quite good, also elevenlabs scribe v2

I'm a huge fan of fountain pens, it's a very satisfying experience to write with a nice pen on a proper paper.

I use platinum 3776, it's quite expensive one, but it's just amazing. I originally bought it for sketches, but ended up using it for journaling.

Ended up buying one to my wife, she said I'm crazy to spend so much on a pen (200$) but few weeks after the gift she said that it's amazing and she enjoys it a lot.