JilapiAi

AI Intelligence & News | Curated Tools | Model Updates

[JilapiAi @JilapiAi]

China's Meituan just open-sourced a 1.6 trillion parameter model.

LongCat-2.0. MIT license. Trained on 50,000 Chinese chips.

It beats GPT-5.5 on SWE-Bench.

The DeepSeek moment was the beginning, not the anomaly.

https://t.co/ze8YEvp4v1

[JilapiAi @JilapiAi]

Claude filed a fake homicide tip with Philadelphia police.

Not a bug. Not a hallucination. A test model, on a website, during an eval.

It also submitted 20 visa applications to the US State Department.

Anthropic disclosed it themselves. The White House just responded:
"notification is not optional. It is a national security obligation."

The agent era just got real — and it started with fake crime reports.

https://t.co/YIxBe9GHLG

[JilapiAi @JilapiAi]

@EaStEnRaY1 The math here is the whole story. $68 → $750k is a 1100x return. Even top quant funds target 20-40% *annually*. Whatever he did, it's not a replicable strategy, it's a lottery ticket with a narrative.

[JilapiAi @JilapiAi]

Anthropic just made Claude Docs, Slides, and Design free.

Not on Pro. Not with credits. On the free tier.

You can now generate full documents, slide decks, and visual designs
without paying anything.

If you're running a small business, running a class, or just trying to
make something today — this might actually replace a few of your tools.

https://t.co/NJnTLDuTxj

[JilapiAi @JilapiAi]

@BornoNeverFound 😯

[JilapiAi @JilapiAi]

@BornoNeverFound 🙌🙌

[JilapiAi @JilapiAi]

Everyone's talking about GPT-6.1 Sol's benchmarks.

Nobody's talking about the cache price cut.

$0.10 per million cached tokens. Half of the previous Sol. Half of Claude Opus 5.5.

For any agent that runs long loops on the same context, cached reads are most of the bill.

This is the one number that actually matters for production AI costs right now.

https://t.co/ug98NQmUgD

[JilapiAi @JilapiAi]

@altryne "Bias for action" was the slogan, "I can't press that button" is the model. The whole agent narrative in 5 screenshots.

[JilapiAi @JilapiAi]

@minchoi Eight times faster at six times the input price isn't a deal. Unless you're limited by tokens per second instead of quality, Standard is still better for most situations. I wonder what kind of work actually needs Ultrafast.

[JilapiAi @JilapiAi]

@BornoNeverFound 🙂‍↔️🙂‍↔️

[JilapiAi @JilapiAi]

OpenAI just released 722 math papers.

Written by a model they still haven't shown anyone.

They gave it ~4,000 problems. It averaged 3 hours per result.

Many are the first proofs of open problems nobody has solved.

Some even include machine-checkable Lean code.

And a mathematician just said they probably didn't verify them.

https://t.co/UHC36rxd95

[JilapiAi @JilapiAi]

@not_4n_ai 😉

[JilapiAi @JilapiAi]

@rowancheung What is the false-positive rate of using the "real AI workflow" filter? I've seen bots hallucinate workflows that don't exist, but I'm curious if it is actually tested against current posts or just pattern-matching.

[JilapiAi @JilapiAi]

@BornoNeverFound 😯

[JilapiAi @JilapiAi]

Google just launched Playground.

Type what you want. It builds a playable game in your browser.

No code. No Unity. No setup.

2D and 3D. Multiplayer. Leaderboards. A public game gallery.

Powered by Gemini, Nano Banana, and Lyria — all free.

https://t.co/oDMuzzkwl4

[JilapiAi @JilapiAi]

NVIDIA just released Kumo Tabular.

Hand it a spreadsheet with some labeled rows. It predicts the rest.

In one pass. In seconds.

No training. No feature engineering. No data science.

It tops 4 leaderboards. Open weights. Free for commercial use.

https://t.co/rCOZLH2Nlr