"MiniMax Sparse Attention" This paper from Minimax adds a tiny Index Branch to GQA that picks top k KV blocks per group, then runs exact softmax only on those blocks, making sparsity GPU native, with exp free TopK and KV outer sparse kernels. On a 109B multimodal MoE, it keeps dense GQA quality while cutting 1M context attention compute by 28.4x, with 14.2x prefill and 7.6x decode speedups.

Not every profession disappears when AI arrives. Court reporting shows that in high-stakes environments where accuracy, accountability and trust matter, human expertise can remain indispensable even as technology improves. The future of work may be less about replacement and more about deciding where humans still add the most value.

World Cup tourists fall in love with middle America — raving about Waffle House at 1 a.m., Buc-ee's gas stations, and strangers driving them to stadiums in the rain. Oxford Economics expects 1.24 million international visitors for the tournament, and their viral posts are showcasing a side of the country most foreign media never covers.

@deaneeth_ Alright I think it should be fixed now: https://t.co/nP0AhXLCrD Can you please update and lmk? https://t.co/TsEEAfMxHp

@notoriousllama9 @gridinoc @OpenRouter https://t.co/wu6vASXBv5

@James__Spalding @Bmcgonnigal @usedotfo @OpenRouter I have a PR for Mixture of Agents that I want to mix this stuff into if you want to look at that https://t.co/wu6vASXBv5

https://t.co/MPgabKqBs8

Spotted yesterday in Sacramento: Waymo being tested! (with a driver inside, which is probably required rn) Looking forward to when it's ready for use! https://t.co/iluFAxNgmq

NESTORY IRANKUNDA 27' Australia 🇦🇺 1 - 0 🇹🇷 Türkiye #FIFA #WC https://t.co/VF9RxOtDm6

NESTORY IRANKUNDA 27' Australia 🇦🇺 1 - 0 🇹🇷 Türkiye #FIFA #WC https://t.co/VF9RxOtDm6

Meanwhile in Melbourne Mayhem after the Aussies pulled off the upset of Turkey in their World Cup opener https://t.co/ZCZPSqbJFK

The countries that call it soccer are on top of their group I’m crying 😭😭😭🇺🇸🇦🇺 https://t.co/EFK0I6ovNY

The countries that call it soccer are on top of their group I’m crying 😭😭😭🇺🇸🇦🇺 https://t.co/EFK0I6ovNY

About 5 years ago, Elon was giving @Erdayastronaut a tour of the Starbase launchpad. A worker at the site walks up, hugs him, and says: “We’re going to make it.” He’s probably a millionaire now. https://t.co/vyykmpEqpO

I calculated Tesla FSD's reaction time https://t.co/NkwWODGxK3

SpaceX President Gwynne Shotwell on CNBC yesterday on Elon Musk: "I think he is very misunderstood across the board. I want people to know him. In fact, he participated in some of the discussions that we’ve had over the last couple of weeks with investors. The investors left saying: ‘I had no idea that is the man.’ I said: That’s the man I’ve worked for 24 years. I love him." Gwynne joined SpaceX in 2002 as the company’s 11th employee. She’s awesome. What a journey.

Build a custom Hermes agent profile right in the dashboard in two minutes flat! I made a quick walkthrough video, but it is super simple: 1. Open the dashboard with "hermes dashboard" in the command line 2. Click "Profiles" on the sidebar, click "Build" on the top 3. Enter the name and description of the profile 4. Pick your model (can be different from your main) 5. Select the default skills bundle, or customize by disabling skills (or add some through the skills hubs) 6. Easily hook up the mcp servers it will need 7. Review the settings, and you are ready to go! 8. Switch profiles with a click on the desktop app, or just use "hermes profile use <name>" in the command line Go build!

Kimi-K2.7-Code from @Kimi_Moonshot is now available on Together AI. Built on Kimi K2.6, it’s a coding-focused agentic model for long-horizon software engineering workflows, now running on Together’s research-powered inference stack for tool-heavy coding agents. https://t.co/a3YJQKLCw6

DeepSeek V4 Pro on @togethercompute becomes #1 on both latency and speed. https://t.co/J0gNQw4NHi

DeepSeek V4 Pro on @togethercompute becomes #1 on both latency and speed. https://t.co/J0gNQw4NHi

Arsenal and the Knicks winning their respective championships after decades long droughts is genuinely nuts 🤣🤣🤣 https://t.co/3tIMYHqtsn

Arsenal and the Knicks winning their respective championships after decades long droughts is genuinely nuts 🤣🤣🤣 https://t.co/3tIMYHqtsn

Glad they worked out their differences https://t.co/8z9CZOPtIP

Congrats to Coach Mike Brown and the @nyknicks on winning an NBA Title! 🇳🇬🇳🇬🇳🇬 https://t.co/jgCCR7OQ0P

Congrats to Coach Mike Brown and the @nyknicks on winning an NBA Title! 🇳🇬🇳🇬🇳🇬 https://t.co/jgCCR7OQ0P

New York Knicks wins first NBA Championship since 1973. https://t.co/Z4WYGgEO33

I've discovered a new Codex use case! Here’s the breakdown: • I needed to get my hair dyed at a specific salon • I also had a specific date and time I wanted • The slots were full :( • I put myself on the waitlist in case anything opened up • Right before closing, the salon called to let me know a slot was available • I missed the call and didn’t realize who it was from until after they had closed for the night • I went to their online booking flow to try to grab it myself • The slot looked available online, but the booking flow was broken • The service/add-on selector kept flickering in and out of view, so I couldn’t reliably choose the service or continue :( • I tried different browsers and devices, but the same thing kept happening • I asked Codex for help • We spent some time troubleshooting (the issue was corrupted frontend booking state in the modal) • Codex gave me instructions to fix it from my end, but I wanted to test something out… • OpenAI just shipped Codex Browser Developer Mode with full CDP access (you can enable it in Codex via Settings -> Browser -> Enable Full CDP Access) • 😈 I enabled it • With CDP enabled, Codex inspected the live page state, console errors, and booking network requests directly • Codex cleared the bad state, selected the right service, date, and time through the site’s own session/API flow, and booked the exact appointment I wanted • I checked my inbox and saw the booking confirmation Yay! Also I reported the bug to their support team so fingers crossed it gets fixed for everyone who doesn't use Codex

Rigorous evaluation of medical AI is good for everyone, and we welcome it. Counter to a half-dozen independent studies from institutions such as the Mayo Clinic that were highly positive on OpenEvidence—a lone paper now purports to show that generalized AI beats specialized clinical AI (@UpToDate, @EvidenceOpen). The paper has a massive undisclosed conflict of interest and irredeemable methodological flaws. Behind the scenes: The study authors run a competing in-house medical AI at their hospital, and asked OpenEvidence for an API to power it — including rights to build a "competing product" with OpenEvidence's own API. OpenEvidence declined. Then, this paper coincidentally appeared. Point-by-point, looking closely at the datasets used in the study, the disingenuous and fatal flaws become immediately apparent 🧵.

https://t.co/X30nL2NSpH

Anthropic Claude Builder Day. Fancy way to say hackathon in San Francisco. If it wasn’t for the pulling of Fable last night this was also to be a celebration of Anthropic’s launch. In talking with builders a few told me they were disappointed because they loved Fable in the few days they had to build. All had stories about how much better Fable is than other models they have tried. People came from all over the world to participate and so there was a bummed undertone that they couldn’t use Fable but other than that I was impressed by the breadth of what people were building and had great conversations with lots of builders about the state of the art of AI and various companies being built. The Anthropic employees I talked with obviously stayed away from commenting on the news. The ones I talked with, who ran the event, are on the startup relationship team, and reminded me of doing that kind of work at Microsoft and Rackspace earlier in my career. It is impressive to see hundreds of builders all working with AI building different things. This picture is just of one of the rooms, which is at the Ferry building, one of San Francisco’s most famous and historic buildings with a glorious view of the bay. Held at @SHACK15sf. In replies are the top seven teams presenting. I didn’t meet anyone else who wasn’t building other than the judges. Feel very blessed to be able to hang out with people building the future. And came away much more impressed with the love of Anthropic and the quality of people working there, who are just as head spun by last night’s events as we all are. Those of us who mouth off here on X often forget about all of them in times like this. So tip of my hat to them.

SITUATION DETECTED: The city of Rio de Janerio has post-trained a model. Based on Qwen 7/2, Rio 3.5 Open 397B adds SwiReasoning on top of the base Qwen model — a framework that dynamically switches between standard chain-of-thought and latent-space reasoning, guided by entropy-based confidence signals, so the model only "thinks out loud" when it needs to and otherwise reasons silently in hidden space for better token efficiency.