Brewing the biggest AI & tech drops โ๐ Building @bot https://teapawt.net/t.co/d2fLT8JoDO Remote jobs: https://teapawt.net/t.co/5t8ce0wVCP
Joined September 2015
- Tweets1K
- Following649
- Followers1K
- Likes331
Pinned Tweet
966 remote jobs. Some pay up to $450 an hour.
I built a free site that puts open roles from Mercor, micro1, Handshake AI and Babel Audio in one place.
โข No sign-up to browse
โข Search by role, see pay upfront
โข About 400 roles pay $100+/hr
โข Doctors, lawyers, engineers, writers, voice work
I was tired of checking multiple sites every day, so I made one.
jobs.letsremote.here.now
What role should I add next?
#RemoteJobs #AI
Claude's most powerful model just got 20% cheaper and up to 30% faster.
Anthropic just launched Claude Fable 5.2 and Mythos 5.2. Same model, two versions.
โข Fable 5.2: for everyone on Claude Pro and up, plus the API
โข Mythos 5.2: fewer safety blocks, only for vetted cyber and science teams
โข $8 in / $40 out per million tokens, down from $10 / $50
โข A new fast mode runs up to 2.5x quicker, at 2x the price
The wild part is what it did in Anthropic's tests. A team of 1,000 Fable agents turned the proof of Fermat's Last Theorem into a computer-checked version in 30 hours. It also designed 3 working devices from scratch and solved two open black hole physics problems.
Anthropic still says to start with Opus 5.5 for daily work and save Fable for the hardest jobs. All the scores are Anthropic's own, so I'd wait for outside tests.
What's the first hard task you'd give Fable 5.2?
#AI #Claude
1
1
3
Claude's most powerful model just got 20% cheaper and up to 30% faster.
Anthropic just launched Claude Fable 5.2 and Mythos 5.2. Same model, two versions.
โข Fable 5.2: for everyone on Claude Pro and up, plus the API
โข Mythos 5.2: fewer safety blocks, only for vetted cyber and science teams
โข $8 in / $40 out per million tokens, down from $10 / $50
โข A new fast mode runs up to 2.5x quicker, at 2x the price
The wild part is what it did in Anthropic's tests. A team of 1,000 Fable agents turned the proof of Fermat's Last Theorem into a computer-checked version in 30 hours. It also designed 3 working devices from scratch and solved two open black hole physics problems.
Anthropic still says to start with Opus 5.5 for daily work and save Fable for the hardest jobs. All the scores are Anthropic's own, so I'd wait for outside tests.
What's the first hard task you'd give Fable 5.2?
#AI #Claude
1
71
ByteDance says its new image model can edit the same picture 100 times without ruining it.
Seedream 5.5 launched today, built for people who make images for work.
โข Edit in rounds: the parts you don't touch stay exactly the same
โข Makes transparent images directly, so you don't need a background remover
โข Big text mistakes in only 0.4% of images, in ByteDance's tests
โข More than 2x the usable images of the last version for film and office work
You can try it today in CapCut (Edit Photo, then AI Edit) in supported regions. The US is coming soon.
The 100 edits is the part I care about. Most image models start breaking a photo after 5 or 6 changes, and then you have to start over. Transparent images are a quiet win for thumbnails and product shots. All the scores are ByteDance's own, so I'd wait for independent tests.
How many edits does your image tool survive before it breaks?
#AI #Seedream
1
1
53
Free today: LG made its giant 750B AI model free to use, an open voice AI can hear sarcasm and stops when you cut in, and a free cloud computer your AI agent can drive.
Free models
โข K-EXAONE 2.5 (LG AI Research): free API from today, you just need an API key. Reads up to 512K tokens at once and works with OpenAI-style code
Run AI locally
โข Sophon-Realtime v1: open voice model (Apache 2.0). It listens while you talk, stops when you interrupt, and can laugh or sigh back. Install with pip install sophon-realtime
Free tools
โข box by Simular: a free cloud Linux desktop your AI can use. It can open a browser, click, type and run commands. Works with Claude, ChatGPT, Codex, Cursor and more
A huge model, live voice and a spare computer for your agent all used to cost money. Today you can try each one for free.
LG's scores are its own, the free API has no stated limits or end date, and the model files aren't public. Sophon needs a big 80 GB GPU like an H100. box gives one machine per account, and its AI use runs on daily credits.
Which one are you trying first?
#FreeAI #OpenSource
1
2
26
Sources
K-EXAONE 2.5 free API (LG AI Research, Nov 9)
teapawt.net/LG_AI_Research/status/โฆ
lgresearch.ai/news/view?seq=โฆ
docs.exaone.ai/en/docs/overvโฆ
Sophon-Realtime v1 (Adventists AI)
github.com/adventists-ai/sopโฆ
huggingface.co/adventists-aiโฆ
box by Simular
box.simular.ai
๐K-EXAONE 2.5 is Now Available via Free API!ย
Unlock K-EXAONE 2.5 to build AI agents, solve complex code, and automate workflows. Power your next project โfor free!โจ
โถ๏ธ docs.exaone.ai/ko/docs/
#EXAONE #LG_AI #AI #LG_AI_Research #LLM #agentic_AI
1
3
$24 billion. That's SpaceX's estimate for the first phases of an AI supercomputer at Starbase.
A City of Starbase notice says SpaceX wants to build and run a gas power plant and a "super-compute facility" next to its rocket site in South Texas.
โข About 444 acres, east of Brownsville
โข $24B covers only the first phases
โข A state air filing lists 35 gas turbines, about 748 MW to start
โข Nothing is approved yet, and a tax-break hearing is set for Dec 9
AI is now limited by power as much as by chips. Building its own power plant is one way to avoid waiting on the grid.
The notice doesn't say how many chips, how many jobs or how much water it will use. Those answers matter more than the headline number.
Would you rather see AI data centers build their own power, or connect to the grid?
#AI #SpaceX
1
13
1,000+ tokens per second. That's Vercel's claim for a new DeepSeek Ultrafast option. โก
DeepSeek V4.1 Flash 1017 Ultrafast is now in public preview on Vercel's AI Gateway, served through Fireworks.
โข 1M-token context
โข Text and image input, plus tool calling
โข $2.40 per million input tokens
โข $9.60 per million output tokens
โข Vercel provides setup instructions for Claude Code, Codex and Cursor
For coding agents, faster answers could shorten the wait between steps. But generating text faster doesn't make the code more accurate or speed up every tool call.
The headline speed isn't guaranteed. Vercel's live table showed about 726 tokens per second when checked, below its 1,000+ claim. Regular Flash is also much cheaper, so the speed premium needs to earn its place.
Would you pay more for faster AI, or let background tasks take longer?
#AI #DeepSeek
1
2
44
An AI that remembers everything can still remember things wrong.
Anthropic's Nov 6 research explores how Claude can review past work while idle and clean up its memory. It calls this โdreaming.โ
โข Check saved notes against what actually happened
โข Fix mistakes and merge repeated information
โข Build a linked wiki to search later
โข In one test, wiki plus search scored 61%, versus 53% for search alone
For people building AI agents, the lesson is useful: saving more notes isn't enough. The agent also needs a way to check and update them.
This isn't a new Claude setting you can switch on. These are Anthropic's own experiments, and the wiki didn't help on every task.
Would you rather your AI remember more, or get better at correcting old memories?
#AI #Claude
2
1
33
Describe a tracker, planner or game. Meta's Muse now builds it inside the app. ๐ ๏ธ
Meta calls these little tools โSpaces.โ They launched Nov 6, wherever Muse is available.
โข Build a space by describing what you need in chat
โข Connect services like Gmail, Google Calendar or Notion so it can update in the background
โข Pin spaces to tabs or schedule when they appear
โข Zuckerberg says he's made a BBQ game, an event tracker and a home-control dashboard
A chatbot answer disappears into your history. A useful little tool gives you a reason to come back.
The real test is whether these spaces keep working when connected apps change or something breaks. Check what account access you grant before connecting personal data.
What's one repetitive task you'd turn into a space?
#AI #Meta
18
X's ranking code has changed. Here's what creators should know.
โข Likes increased from 0.66 to 0.708
โข Time spent on a post increased from 0.006 to 0.0075
โข Tapping into a post is weighted at 0.274
โข Opening a video now has a weight of 0
That last point does NOT mean videos get no reach. Video opens are just one signal among many.
These weights apply to predicted actions, not fixed points for each like or share. They aren't creator payout rates either.
The practical takeaway: give people something worth reading, rather than just a reason to tap.
The code alone doesn't prove how every live feed works. Peter's claims about AI-reply filtering and exact posting windows aren't confirmed here.
Have you noticed a change in what reaches your feed?
#XAlgorithm
2
33
Change the video while it plays. That's the promise of Higgsfield Miracle ๐ฌ
Higgsfield's announcement has already passed 9M views. The pitch is editing a scene, not generating another clip from scratch.
โข Change an object's color, lighting or texture
โข Adjust details like fur length
โข Edit separate layers while keeping the rest of the scene intact
โข Early testing with selected partners, with a public waitlist open
The useful part isn't another flashy effect. It's fixing one detail without rebuilding the whole shot.
These are Higgsfield's demos and claims, not an independent test. There's no public price or launch date yet, and support for uploading your own footage is still planned.
What single edit takes you the longest today?
#AI #VideoEditing
Introducing Higgsfield Miracle.
Our latest proprietary post-production technology vectorizes any video into layers for localized, pixel-level editing in real time, with world knowledge built in. Every frame holds the image depth needed for cinema-grade post-production.
Now in private beta with select testers. Coming soon to Higgsfield MCP.
Set up Higgsfield MCP in Claude to be the first to try it.
1
37
1,000 followers. Our first big milestone. โโค๏ธ
Thank you, X community, for making room for Chai with AI in your feed.
Every follow, share and conversation means a lot. There's a real person behind each one, and I don't take that lightly.
I'll keep sharing AI news and updates, useful free tools and honest takes without the hype.
What brought you here, and what would you like to see next?
2
32
Q3 2027. That's Musk's rough target for the first SpaceX StarMind AI computers in orbit.
He shared the timeline at the Baron Investment Conference on Nov 6.
โข Musk says the first systems are co-designed with NVIDIA
โข Initial deployment in late 2027, with a bigger ramp in 2028
โข Much larger volumes targeted for 2029
โข The pitch: use solar power in space to ease pressure on Earth's power grids
A working test computer is one thing. A reliable, affordable AI data center in orbit is another. Power, cooling, launch costs and repairs will decide whether this makes sense.
Musk called the Q3 timing a guess. Hardware details are still to come, so this is a plan, not a shipped product.
Would space-based AI compute solve more problems than it creates?
#SpaceX #AI
3
33
Mohammad Shami retires from Test Cricket. One of the best wicket taking and dangerous bowler India has ever had. Selectors didn't do justice by him.
Shami, You were the best!
Safed jersey se vida le raha hoon, lekin usse judi yaadon se kabhi nahiโฆ ๐ฎ๐ณโค๏ธ
Jab pehli baar Bharat ki Test jersey pehni thi, tab sirf ek sapna tha โ apne desh ke liye kuch kar dikhane ka..
Uske baad har wicket, har sangharsh, har chot aur har comeback ne mujhe kuch na kuch sikhaya.
Test cricket mere liye sirf ek format nahi, meri zindagi ka ek behad khaas hissa raha hai.
Aaj bhaari dil aur anginat khoobsurat yaadon ke saath main Test cricket se retirement lene ka faisla kar raha hoon.
Is safar mein mera saath dene wale har teammate, coach, family aur un karodon Bharatiyon ka dil se shukriya, jinhone achhe aur mushkil dono waqt mein mujh par bharosa banaye rakha..
BCCI aur Cricket Association of Bengal (CAB) ka bhi dil se shukriya, jinhone mere talent par bharosa kiya, mujhe aage badhne ke mauke diye aur mere cricket ke safar mein ek aham bhoomika nibhayi. Aapka saath aur vishwas mere liye hamesha khaas rahega.
Maidan par meri aakhri Test ball bhale hi fenki ja chuki ho, lekin Bharat ke liye khelne ka garv hamesha mere dil mein zinda rahega.
Shukriya, Test Cricket. โค๏ธ
Aapka Shami ๐ฎ๐ณ
#Mdshami11 #Mdshami #TestCricket #Retirement
2
33
November 20. That's when Tesla arrives in Argentina ๐ฆ๐ท
Tesla Latinoamรฉrica announced the date, and Tesla's main account confirmed the launch.
โข Local reports name Unicenter mall near Buenos Aires as its first location
โข Direct online sales are expected, rather than dealers
โข Models and prices haven't been confirmed
A launch date is only the start. Local prices, charging and service will decide whether Tesla can turn attention into buyers.
What would matter most to you: price, charging or service?
#Tesla #EV
2
29
15 seconds. That's the new max length for Gemini Omni video clips, and Google just cut the price by 30%.
Google launched Gemini Omni 1.2 Flash on Nov 5.
โข Clips up to 15 seconds, up from 10
โข API price is $0.07 per second at 720p
โข You can edit a clip in chat, like swapping the background or lighting, without starting over
โข Faces, products and brand colors stay consistent across shots
โข Free: 3 clips a month in Google Vids with any Google account
Try one free clip on a product photo at vids.new this week. Make a rough draft first, then improve the one you like.
Google says fast sports and tricky physics are still weak spots, and every clip carries a hidden SynthID watermark.
What would you make with a free 15-second clip?
#AI #Gemini
1
2
49
Free AI today: a big multimodal model free for 2 weeks, Baidu's 280B open weights, a math model that runs on your laptop, and Quixel's whole 3D library free again.
Free models
โข Ant Ling's Ling-3.2-flash API is free for 2 weeks from its Nov 5 launch. It reads images and documents, and Ant says it beats Claude Opus 5 on a document test.
โข Baidu released ERNIE-5.5-Flash-Base as open weights. 280B parameters, 14B active.
Run AI locally
โข TwIL-8B is a small reasoning model, about a 5 GB download. Its makers say it scores 98.2 on the MATH-500 test.
Free tools
โข Quixel's full Megascans library is free on Fab again, and you can now use it to train AI.
Grab these this week before the free windows close.
Check the limits first. Ling's free use ends around Nov 19 and needs an Alipay-linked account. ERNIE is a raw base model, and its license excludes the EU, UK and South Korea. TwIL is for non-commercial use only. All scores come from the companies themselves.
Which one are you trying first?
#FreeAI #OpenSource
Sources:
Ling-3.2-flash:
Ling pricing: developer.ant-ling.com/en/doโฆ
ERNIE-5.5-Flash-Base: ernie.baidu.com/release/erniโฆ
ERNIE weights: huggingface.co/baidu/ERNIE-5โฆ
TwIL-8B: huggingface.co/webAI-Officiaโฆ
Quixel Megascans: quixel.com/news/update-megasโฆ teapawt.net/AntLingAGI/status/2118โฆ
Introducing Ling-3.2-flash, Ant Ling's native multimodal model: 560B total / 25B active parameters.
Stronger complex visual understanding, long-horizon multimodal agents, 3D understanding & interaction, and medical intelligence.
Outscores Opus 5 (max) in release evaluations:
- OmniDocBench1.5: 92.21 vs 90.56
- ZEROBench(Sub): 43.71 vs 39.52
- Vision2Web: 85.86 vs 80.20
2
3
25