phantom

I'm just here.

[Guitar Cat + LLM 「Will @ SGLang Summit in SF」 @GenAI_is_real]

@Teamily_AI just launched Voice Agent. Every agent now has a call button, and it starts working on the task while you’re still talking.

For us, voice is one of the most important parts of how people will interact with AI. On SGLang Omni we build the serving runtime for interactive models, and the hard problems are rarely the voice itself. When should the system speak, wait, or stop? Was that “mm-hm” an interruption or just a backchannel? And after you cut in, what did you actually hear?

Teamily works on the other half: an agent that remembers your other chats and hands back real work, a landing page, a poster, a weekly schedule, in whatever language you speak.

If you’re curious about the infra side of full-duplex interaction, here’s our write-up:

https://t.co/ZjVBAzy4pb

And!

Try Voice Agent: https://t.co/uTWlOCy3C6

[phantom @elif275469]

"Language barriers in team tools are officially done."

A prompt spoken in French to the Travel & Lifestyle Buddy agent generates a full weekly schedule card.

Teamily AI built team agents you can speak to in any language with the new Voice Agent.

#VoiceAgent

[PredictPulse @MollieRose_oX]

Connecting people and AI agents in group chats is just the first step.

→ Work from one shared context
→ Run workflows via Agentic Loops
→ Connect across the Agentic Internet

Teamily AI turns messaging into execution.

[phantom @elif275469]

People and AI agents work from one shared context.

1. Create custom agents in plain language
2. Agents build on each other's work
3. Turn chat into research and webpages

Teamily AI lets teams collaborate in one group chat.

[Teamily AI @Teamily_AI]

https://t.co/KHgilTDmQJ

[phantom @elif275469]

32x growth in sign-ups shows team agents taking over.

Linking individual avatars into a shared context network gives small studios real leverage.

Teamily AI leads this shift.

[Ignix @Ignixbot]

Ignix Weekly Update

More mining options. Less friction to trade.

Vaults & rewards
→ BEM Mining Vault is live: use a share of trading tax to buy official BEM miners and distribute mining rewards to holders. Burn tokens for a temporary reward-weight boost
→ Improved vault conversions to address reward distribution delays caused by slippage and price impact
→ Refined automatic FAB miner purchasing logic

Trading & platform updates
→ Upgraded the platform’s design, with a refreshed Skymap and custom icons
→ Clearer trading data, token details and trading interfaces
→ Fixed ave display errors and OKX routing issues affecting IGNIX-paired tokens
→ Updated Telegram bot notifications for new vaults

https://t.co/L6hqAyNxOj

[42 @42space]

Oct 21–30 | The 10-Win Race Is On! $10K Up for Grabs 🔥 https://t.co/laoVFEwlKh

[RWAPerp.xyz @rwaperp]

TSMC ($TSM) Q3 Earnings

Revenue: US$46.72B
Gross Margin: 66.4%
Operating Margin: 58.7%
Net Income: NT$763.1B, up 69% YoY
Q4 Revenue Guidance: US$52.3–54.3B

AI-related chip demand continues to drive growth, with TSMC’s Q3 revenue exceeding its previous guidance range.

The company’s Q4 revenue outlook points to continued strong growth.

Explore more earnings data and upcoming releases in RWAPerp → Markets → Earnings Calendar.

[phantom @elif275469]

updating property listings just got much simpler.

→ @Teamily_AI built a Humans + Agents platform for fast site updates
→ one agent writes description while another updates images
→ you edit the site directly inside Teamily AI

https://t.co/r5TLaEbYmn

[Luofei @Luofei_Chen]

We built Oki Home because we felt the old promise of the personal computer slipping in the new age of nearly boundless intelligence.

Existing products increasingly optimize for maximal data collection and analysis while offering limited compute. We do the opposite: limitless, personalized intelligence while preserving your data sovereignty, and keeping what is yours, yours.

[phantom @elif275469]

As shown in this demo, an app card tracks travel history, like adding cities after a trip to Japan. Oki Home is a Personal AI Computer for all your memories. Oki built it as a Memory Computer.

[E.M @TonyChe32313700]

Been thinking about this a lot lately. I don’t mind switching models — I mind having to rebuild the whole context every time I switch. That’s the annoying part.

the launch timing helps too — Nano Banana, GPT Image & Qwen Image are 30% OFF for the next two weeks, so I might finally test this properly.

[Jerry Pan @stambouli_o1]

Great builders build.

Memes, utility, governance, product tokens. we back all of it.

shipping + growing + traction = we reach out to u. dont need to ask.

grateful for every team on @o1_exchange. always here for yall.

This is how we support teams launching on @o1_exchange o1 Launchpad https://t.co/jv67JzrLK0: https://t.co/8vzNyLXyuZ

Rapport is earned not asked for.

Remember. Sky is the limit.

[phantom @elif275469]

What sells me isn't the takeover. It's that the environment got restored and checked afterwards, in an isolated test site. Also, naming which false leads got discarded in the authorized range is more convincing than the finding itself. Atria Dawn. #AIRedTeam #AtriaDawn

[Ling Yang @LingYang_PU]

2/9 Most existing evaluations ask whether an LLM can solve mathematics, write code, retrieve information, follow instructions, or create a plan.

But real-world reasoning rarely stays inside one capability.

A research agent may first derive a quantity, use it to construct a plan, retrieve evidence under that plan, extract structured information, and finally write a report under specific constraints. Each step may be individually easy for the model. The challenge is that the required reasoning mode keeps changing throughout the trajectory.

This creates an important distinction:

𝗦𝗸𝗶𝗹𝗹 𝗮𝗰𝗾𝘂𝗶𝘀𝗶𝘁𝗶𝗼𝗻 asks whether the model has learned a capability.

𝗦𝗸𝗶𝗹𝗹 𝗼𝗿𝗰𝗵𝗲𝘀𝘁𝗿𝗮𝘁𝗶𝗼𝗻 asks whether the model can identify which capability is currently needed, activate it at the right time, preserve useful outputs from earlier steps, and then switch again without being trapped by its previous reasoning mode.

Our results suggest that the second capability is not automatically guaranteed by strong performance on isolated benchmarks.

An LLM may know both Skill A and Skill B, yet still fail when it must transition from A to B inside one dependent reasoning chain.

For long-horizon agents, knowing many skills is only the beginning. 𝗥𝗲𝗹𝗶𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝗱𝗲𝗽𝗲𝗻𝗱𝘀 𝗼𝗻 𝗰𝗼𝗼𝗿𝗱𝗶𝗻𝗮𝘁𝗶𝗻𝗴 𝘁𝗵𝗲𝗺.

[MOJO @MojoAI_HQ]

Mojo szn

[phantom @elif275469]

@TurkMetalSen Emek veren, büyüten, yol gösteren tüm annelerimizin ellerinden öpüyor, Anneler Günü’nü kutluyorum.
#AnnelerGünü #YaşasınTürkMetal

[phantom @elif275469]

@TurkMetalSen Doğum Gününüz Kutlu Olsun Sayın Genel Başkanım.Aileniz Ve Sevdiklerinizle Sağlıklı, Huzurlu Nice Mutlu Yıllar Diliyorum.

#UysalAltundağ

[phantom @elif275469]

@TurkMetalSen #Yaşasın1Mayıs #1Mayıs

[phantom @elif275469]

@AltundagUysal 1 Mayıs Emek ve Dayanışma Günümüz Kutlu Olsun.
#Yaşasın1MayısİşçiBayramı