Trion

@TheMathFlow Really liked this, great explaination.

The Geometry of Integration by Parts.

@Voxyz_ai Sad that it doesn't cover models like Minimax M3, Deepseek, MiMo, Qwen etc.

feels like this is the only leaderboard you need for ai coding. one thing though, read cost and steps together, otherwise you can't make sense of it.

@Angaisb_ It has 2x usage as of now. You get 60$ of usage in 5$ price, so it is way worth it. Kimi K3 requires more tokens than other flagships and costs 15$/M tokens so you can only expect ~10 usages. I personally keep it for MiMo 2.5 model for coding tasks, and GPT for planning.

Anyone here with an OpenCode Go plan? Does it offer enough Kimi K3 usage? It's just $5 for the first month so I'm tempted

@thsottiaux Being able to make music/songs using GPT?

How can we make ChatGPT better. It’s already on your phone, but what do you wish would happen more or less of when you open the app. How can it bring you more joy?

@uzairansar Now there is 2x usage on Opencode Go, you can double demolish it!

I am about to demolish my OpenCode usage with Kimi K3

@aleksliving I think that is ADHD. Do you start to learn and then you get distracted in few days and now you believe you can learn another skill in 2 weeks? Cycle goes on...

what’s the name of the mental illness where you think you can learn literally anything in two weeks? just got diagnosed

Meh, I didn't even use Grok 4.5 that hard. Quality and quantity of work, both were worse here.

Tried Grok 4.5 and it just seems to need a lot of course correction compared to GPT 5.6 Terra or Luna. Not sure what's missing.

@birdabo Make a 3d animated scene from Matrix

what should i ask first? kimi credits in replies.

@shaunralston @thsottiaux Didn't recieve it 🙃

thx @thsottiaux 🚀🔥

@argofowl Really liking the idea of pace visibility. That would allow user to know that they need to decrease the effort if they do not want to deplete usage so quickly. Currently with invisible token/price limit, we have to guess how fast we are going based on percentage per minute.

here’s what you need to do for the best user experience: - keep the 5h limit gone - add intelligent pace alerts that nudge you when you’re on track to exhaust your weekly usage before it resets, similar to codexbar and openusage - make your pace visible in the ui, and turn it red when you’re going too fast, with the estimated day you’ll hit the limit otherwise people will burn through their usage without realising it, then complain when it’s gone

@ScienceNews Why were landslides surprising though? Is it really such a rare phenomenon for planets?

Pluto is home to landslides, New Horizons imagery reveals. sciencenews.org/article/pluto-…

@haipingfu Interesting results, I would've thought Rust would end up beating Go in all benchmarks. What causes Go to be better at these scans and reads?

Two prolly trees. Two languages. Different strengths. We benchmarked our Rust prolly-tree implementation against the Go implementation from Dolt using identical deterministic workloads at up to 10M records. Both implementations provide content-addressed, persistent ordered maps. The goal isn’t to crown one universal winner—it’s to give technical teams strong Go and Rust options with clearly measured trade-offs. The Rust implementation currently specializes in high-throughput construction, append-heavy ingestion, and bulk mutation. The Go implementation demonstrates excellent traversal performance, leading full range scans and most fresh-tree point-read workloads in this benchmark. That gives us a clear roadmap for the next Rust release: improve point reads and range scans while preserving its write throughput. We’re profiling traversal, decoding, allocation, and cache-locality hot paths now. Methodology: 1. One worker per implementation 2. In-memory storage 3. 10M fixed-width string keys 4. Deterministic 1–100-byte values 5. Identical workload digests and operation counts 6. Validated point-read values, scan ordering, and cardinality The 10M figures are one complete measured pass; three-run tests through 1M confirmed the same overall direction. 10M records — fresh tree construction (All values are ns/op; lower is better): Rust prolly tree: github.com/crabbuild/prol… Go prolly tree: github.com/dolthub/dolt/t…

Antigravity not allowing users to use third party harnesses and actively banning them is probably the worst customer experience in all of LLM subscriptions.

@JJTheHammer29 @vishhvak @thsottiaux @robin_faraj Probably for the time when OpenAI has enough budget and usage is running low

@Trion129 @vishhvak @thsottiaux @robin_faraj Exactly, but what was the point in creating banked resets then

I hit my claude limit, resets in 4h what should I do in the meantime?

@JJTheHammer29 @vishhvak @thsottiaux @robin_faraj With banked resets they have to definately give 100% weekly usage as users will wait to deplete their current quota. With standard/forced reset, they save on usage as most people will still be on 60-100% bracket, so they are effectively only giving 40% reset instead of 100%

@vishhvak @thsottiaux @robin_faraj They’re choosing standard resets over banked ones as it’s less strain on the system somehow

@thsottiaux A guardrail to ask before deleting large scale directories might help.

On file deletions. We’ve investigated a handful of reports where GPT-5.6 unexpectedly deleted files. What we have found is that this most commonly occurs when: - Full access mode is enabled and codex is run without sandboxing protections, including without auto review being enabled - The model attempts to override the $HOME env var to define a temporary directory. - The model makes an honest mistake and mistakenly deletes $HOME instead. This is of course not how we want the system to behave, even when a user operates the model in full-access mode without the safeguards of our sandbox or without using auto review which checks for these kinds of high risk actions and rejects them. We are taking steps to mitigate this risk including by updating the developer message, guiding more users towards safer permission modes, and adding additional harness safeguards. Even though this happens extremely rarely, we’ll share a detailed post-mortem in the coming days that goes into more details and what we are doing to minimize risks further.

@machazthegamer @thsottiaux I believe usage provided is quite dynamic and it was never apple to apple comparison. It is better to compare the raw API costs as model usage can be more or less depending on OpenAI's budgeting for B2C subscriptions.

@thsottiaux Removing 5h limit is sus. You are removing the anchor we had of comparing 5h limit on 5.5 to behavior in 5.6. Now you can squeeze us and shirt chamge

Another reset for our Codex and ChatGPT Work users. Actually hit 9M active users way earlier today, but then got distracted by the approximately millions of things the team is doing to keep the systems up and reliable. Should have that sweet 100% weekly usage limit back in a few minutes. Go be your productive self and close twitter. Shoo!

@thsottiaux Removing the 5 hour limit is better. It would help to allow users to put custom expenditure budgets to keep things in control. But to think of it, isn't effort/juice exactly that?

We've had no 5h limit in Codex plus and pro for a few days. Do you think it is better or are you finding it difficult to manage the usage included in the weekly limit effectively? If we were to make this different, what should it look like in an ideal world?

Hidden cost of local LLMs is the power consumption, the hardware degradation and the lack of long context windows. Kind of makes it worth paying for subscriptions.