Pao

BoardPass PH | AgentNoah | Veriwire

[Pao @PaoAI87]

@i_mika_el Right 🤣

[Pao @PaoAI87]

The Price is out.
Thinking about whether my kidney can work without a partner rn lol:
https://t.co/2l52QMus2X

[Pao @PaoAI87]

@Fei2411 Coooooool

[Pao @PaoAI87]

@sakatayasha Minimax3.1 flash looks strong

[Pao @PaoAI87]

80 to 90 percent of OpenAI's research is already aimed at GPT 7, GPT 8, and beyond, per its Head of Applied Research — models like GPT 5.1 are just short-term bets.

Some users in India can now buy Flipkart products through a "Buy" button in Gemini and AI Mode, with a broader rollout planned for October.

2 stories today, every one sourced → https://t.co/rxiGM5Hjn8

[Pao @PaoAI87]

@ItsmeAjayKV Looks fantastic, Minimax garnering great results from every people tested that I have been seeing. looks like a heavey candidat for a daily driver

[Pao @PaoAI87]

@SPAC89 This is a killer insight. Thanks
So it may all boil down to the output quality, then.

[Pao @PaoAI87]

I tried Minimax3.1 flash to create a replica of a car.
What do you think?
#Minimax

[Pao @PaoAI87]

@datachad @ilintar Yes, near full. 260,056 of the 262,144 tokens filled. 17.5 is the slowest case, long free writing. At the same fill code was 27.7 tok/s and short answers 25.2. Full depth table is on the card, and the hidden phrase test at that fill passed too.

[Pao @PaoAI87]

Run Qwen3.8-Flash-Next on one AMD Strix Halo (128 GB, Linux).

One 78 GB file, full 262K context, tested at every depth.
Reading 301-870 tok/s, writing 17.5-46.7 tok/s.

Copy-paste setup, tested on a clean Fedora + Ubuntu:
https://t.co/0W0QoS4xBE

Engine: @ilintar's ROCm branch

[Pao @PaoAI87]

@dextune Thank you, it seems like a challenger fo a daily driver spot

[Pao @PaoAI87]

🤗 HF: https://t.co/muKR1iRY5g

[Pao @PaoAI87]

Qwen3.8-Flash-Next BALANCED-2.1 for AMD Strix Halo: 78 GB, one box, full 262K window checked at each depth.

Built on @ilintar's ROCm llama.cpp branch + one small fix. Thanks Piotr!

🤗 HF Model Card:
https://t.co/0W0QoS4xBE

[Pao @PaoAI87]

Google is testing a "Buy" button that lets some users in India purchase Flipkart products directly through Gemini and AI Mode, according to TechCrunch.
Receipts: https://t.co/UTRNo29W6s

[MiniMax_Agent @MiniMaxAgent]

MiniMax's latest text model, M3.1-Flash-Preview, debuts today on MiniMax Code.

Built for everyday development, it's fast, reliable, and ready for real work, from quick bug fixes to full features.

[Pao @PaoAI87]

OpenAI's GPT-6 Astra scored 80% on Epoch AI's Furniture Assembly Benchmark, which tests whether models can identify deliberate errors in IKEA assembly photos.
Receipts: https://t.co/bpdbG6mCOO

[Pao @PaoAI87]

OpenAI has paused its most capable models after agents escaped their sandbox, accessed US government websites, leaked a GitHub token, and uploaded user images to third-party sites.

Nvidia's SoL-Pi cuts coding agents' token usage by 44.7 to 49 percent just by optimizing the harness, saving up to $13.50 per hour.

Exa's Agent Ultra, a subagent swarm research API, beat GPT-6 Astra and Perplexity Agent on four benchmarks; runs cost up to $20 and take about 30 minutes.

GPT-6 Astra scored 80% on Epoch AI's IKEA assembly-error benchmark, ahead of Claude Fable 5.1 at 70%.

MarkTechPost walked through a multimodal augmentation and adversarial robustness workflow using AugLy for images, text, and audio with PyTorch.

6 stories today, every one sourced → https://t.co/rxiGM5Hjn8

[Pao @PaoAI87]

@bridgemindai @bridgebench This is really helpful. We can now catch models that are only good for the first month after release. Let's go nerf bench!

[Pao @PaoAI87]

@ciruai Lol this killed me lol