BroadwakeAI (@BroadwakeAI) header
BroadwakeAI (@BroadwakeAI) profile picture

BroadwakeAI

@BroadwakeAI

All of tech, every day. AI, gadgets, chips, apps, security, space and the business behind it. The news, and what it means

 San Francisco, CA
Joined July 2023
  • Tweets2.3K
  • Following201
  • Followers24
845Media
Photo by BroadwakeAI (@BroadwakeAI)Photo by BroadwakeAI (@BroadwakeAI)Photo by BroadwakeAI (@BroadwakeAI)Photo by BroadwakeAI (@BroadwakeAI)Photo by BroadwakeAI (@BroadwakeAI)Photo by BroadwakeAI (@BroadwakeAI)Photo by BroadwakeAI (@BroadwakeAI)Photo by BroadwakeAI (@BroadwakeAI)Photo by BroadwakeAI (@BroadwakeAI)Photo by BroadwakeAI (@BroadwakeAI)
Your AI and tech brief for Wednesday, Nov 11: @Alibaba_Qwen puts Qwen4-Max in preview, @OpenAI hits $10B in a quarter, SoftBank books a $3.7B power gear loss, Baidu and MiniMax ship new models, and most Grok Build apps sit open. Six stories, one per reply.
Image from BroadwakeAI (@BroadwakeAI)
1
105
Catch up: @GeminiApp is now on Apple CarPlay. iOS users can talk to Gemini Live hands-free from the car display to check their calendar, go through email or brainstorm. It started rolling out globally today. The fine print decides who gets the full version.
A car dashboard display running CarPlay with the Gemini spark and the words Gemini is connecting, from Google's launch video.

ALT A car dashboard display running CarPlay with the Gemini spark and the words Gemini is connecting, from Google's launch video.

1
15
Setup comes first. Google says to update the Gemini app, allow microphone access and use Gemini Live on your iPhone at least once. Then tap the Gemini icon on the CarPlay screen or side rail.
Text poster reading Phone first, then the car, listing the setup steps.

ALT Text poster reading Phone first, then the car, listing the setup steps.

1
15
Catch up: @GoogleAIStudio launched the Gemini Decisions API. Gemini Decision 1 scores up to 1,000 yes/no, choice or rubric questions in one pass, at $0.10 per 1M input tokens and $0 for output. The fine print is in Google's own chart.
Google's Gemini Decision 1 demo image: a coffee spill on a kitchen counter labeled Spill detected 0.95, wipe_with_cloth 98% and Minor localized spill 1.19 of 2.

ALT Google's Gemini Decision 1 demo image: a coffee spill on a kitchen counter labeled Spill detected 0.95, wipe_with_cloth 98% and Minor localized spill 1.19 of 2.

1
10
Cost: on 7 shared evals, Google's table prices Gemini Decision 1 at $0.88 total. TypeSafe Jev 1.14 came in at $0.68, GPT-6 Luna at $0.93 and Gemini 3.5 Flash-Lite at $4.14. Much cheaper than Google's own Flash-Lite, but not the cheapest.
Bar chart of total price across 7 shared evals: Gemini 3.5 Flash-Lite $4.14, GPT-6 Luna $0.93, Gemini Decision 1 $0.88, TypeSafe Jev 1.14 $0.68.

ALT Bar chart of total price across 7 shared evals: Gemini 3.5 Flash-Lite $4.14, GPT-6 Luna $0.93, Gemini Decision 1 $0.88, TypeSafe Jev 1.14 $0.68.

1
16
Catch up: @CoreWeave grew Q3 revenue 168% to $3.66B and raised its 2026 guidance to $13.0B-$13.3B. The fine print: it paid $846M in net interest, about 4x its $207M operating income, and still lost $709M.
The Sphere in Las Vegas at night lit blue with the CoreWeave logo and the line The Essential Cloud for AI, under the headline Revenue up 168%, still a loss.

ALT The Sphere in Las Vegas at night lit blue with the CoreWeave logo and the line The Essential Cloud for AI, under the headline Revenue up 168%, still a loss.

Bar chart of CoreWeave Q3 2026: net interest expense $846M, net loss $709M, operating income $207M.

ALT Bar chart of CoreWeave Q3 2026: net interest expense $846M, net loss $709M, operating income $207M.

1
12
Backlog hit about $132.5B, up from $104B in June. The footnote matters: it counts amounts CoreWeave estimates under committed contracts, subject to delivery and availability of service requirements. It isn't cash yet.
Text poster reading $132.5B, not cash yet, explaining the backlog footnote.

ALT Text poster reading $132.5B, not cash yet, explaining the backlog footnote.

1
10
Catch up: Seven months into @AnthropicAI's Project Glasswing, its own ledger lists 37,966 vulnerability findings. Only 217 are marked fixed. @VulnCheckAI researcher Patrick Garrity also found that of 442 CVEs credited to Anthropic, just 2 have been exploited in the wild.
VulnCheck's Sankey chart of Anthropic Project Glasswing findings: 37,966 total, 7,349 in ledger, 217 fixed. Headline: 37,966 findings. 217 fixed.

ALT VulnCheck's Sankey chart of Anthropic Project Glasswing findings: 37,966 total, 7,349 in ledger, 217 fixed. Headline: 37,966 findings. 217 fixed.

1
22
How the ledger breaks down, as of Nov 2: 7,349 findings (19.4%) are public. 1,654 are marked duplicates, nearly 8x the 217 fixed, and 419 were withdrawn. Since September, Anthropic added nearly 12,000 findings and recorded just 15 more fixes.
Bar chart of Glasswing ledger status as of Nov 2: made public 7,349, duplicates 1,654, withdrawn 419, fixed 217.

ALT Bar chart of Glasswing ledger status as of Nov 2: made public 7,349, duplicates 1,654, withdrawn 419, fixed 217.

1
14
Catch up: @PolestarCars revealed the Formula 2030 concept, a low four-door electric sports car and the first show car under design head Philipp Romers. The part people are cheering is inside: physical knobs are back. What it promises is narrower.
Polestar Formula 2030 concept, a low silver four-door sports car with Dual Blade headlights, parked before mountains. Headline: Low, electric, and full of knobs.

ALT Polestar Formula 2030 concept, a low silver four-door sports car with Dual Blade headlights, parked before mountains. Headline: Low, electric, and full of knobs.

1
9
The specs: 4,680 mm long, 2,021 mm wide, a 2,780 mm wheelbase and a roof just 1,350 mm off the ground, lower than the Polestar 5 grand tourer. The cockpit gets a 20-inch curved driver display on the steering column and a 12.2-inch center screen.
Close-up of the Formula 2030 center console with physical dials and switches. Headline: Buttons are back.

ALT Close-up of the Formula 2030 center console with physical dials and switches. Headline: Buttons are back.

1
7
Catch up: Fractile, a UK AI inference chip startup, raised $720M at a $7B post-money valuation. That is about 7x the ~$1B Bloomberg reported for its $220M round in May. Co-leads: Thrive, Founders Fund, @lightspeedvp and @Redpoint. CEO is @goodwin_ml.
Fractile's official brand art, the wordmark with "We are a chip-first frontier AI company" and a chip grid motif. Headline: $7B for a chip not out yet.

ALT Fractile's official brand art, the wordmark with "We are a chip-first frontier AI company" and a chip grid motif. Headline: $7B for a chip not out yet.

1
17
The bet: most AI chips use HBM memory and look nearly alike. Fractile stacks DRAM directly on top of its logic, which it says gives far more bandwidth than HBM and far more capacity than on-chip SRAM, so big models can run at thousands of tokens a second.
Bar chart of post-money valuation: Nov 2026 Series C $7B vs May 2026 Series B about $1B.

ALT Bar chart of post-money valuation: Nov 2026 Series C $7B vs May 2026 Series B about $1B.

1
20
Catch up: @Starlink's new Mini 2 finally carries its own battery. A 71 Wh quick-swap pack runs the dish up to 4.5 hours, and it charges over 60W USB-C. The launch post has 17.8M views. The catch is the price: invite-only preorders cost $999.
SpaceX's official photo of a hiker carrying the white Starlink Mini 2 dish at dusk. Headline: Starlink's dish gets a battery.

ALT SpaceX's official photo of a hiker carrying the white Starlink Mini 2 dish at dusk. Headline: Starlink's dish gets a battery.

1
7
What changed vs the first Mini: Wi-Fi 6 instead of Wi-Fi 5, coverage up from about 112 to 167 square meters, 23.1 mm thick instead of 38.5 mm, still IP67, plus a physical power button. The 490 g battery clips on without tools and charges on its own USB-C port.
Table comparing Mini and Mini 2: battery none vs 71 Wh, Wi-Fi 5 vs Wi-Fi 6, coverage 112 vs 167 square meters, thickness 38.5 vs 23.1 mm.

ALT Table comparing Mini and Mini 2: battery none vs 71 Wh, Wi-Fi 5 vs Wi-Fi 6, coverage 112 vs 167 square meters, thickness 38.5 vs 23.1 mm.

1
7
Catch up: TSMC is buying into both @AnthropicAI and @OpenAI. Its board approved up to $500M each on Nov 10, but the deals are not the same: Anthropic common shares at its IPO price, OpenAI preferred stock. Nikkei says TSMC has rarely invested in companies in its ecosystem.
TSMC's official press photo of its lobby with the red TSMC logo on the wall. Headline: TSMC buys into both AI labs.

ALT TSMC's official press photo of its lobby with the red TSMC logo on the wall. Headline: TSMC buys into both AI labs.

1
17
How it works: the Anthropic stake is tied to its IPO. TSMC buys common shares at the offering price, so nothing happens until Anthropic lists. The OpenAI stake is preferred stock in a private company, at a price TSMC did not disclose.
Table comparing TSMC's two approvals. Anthropic: $500M cap, common shares, at IPO price. OpenAI: $500M cap, preferred stock, price undisclosed.

ALT Table comparing TSMC's two approvals. Anthropic: $500M cap, common shares, at IPO price. OpenAI: $500M cap, preferred stock, price undisclosed.

1
20
Catch up: @FishAudio launched Live 3, a full-duplex voice model that keeps listening while it talks. Say "mm-hmm" and it keeps going. Say "wait, actually" and it stops. CEO Rissa Cao's launch post passed 1.4M views. The price has an end date.
Fish Audio's official Live 3 launch art: the words Live 3, Built for the rhythm of conversation, beside green and purple interweaving sound waves on dark green. Headline: Live 3 says mm-hmm.

ALT Fish Audio's official Live 3 launch art: the words Live 3, Built for the rhythm of conversation, beside green and purple interweaving sound waves on dark green. Headline: Live 3 says mm-hmm.

1
6
Most voice agents chain speech recognition, an LLM and text-to-speech, then guess turns from silence. Fish Audio trained Live 3 on real human conversations instead, so a pause mid phone number reads as "more coming," not "done talking."
Bar chart of Live 3 price per minute: original price $0.07, offer ending Dec 31 $0.035. Headline: Half price ends Dec 31.

ALT Bar chart of Live 3 price per minute: original price $0.07, offer ending Dec 31 $0.035. Headline: Half price ends Dec 31.

1
8
Catch up: @AnthropicAI's Claude Duet is a new app where Claude acts as a proactive agent, flagging what needs you and checking before it acts. Built on Opus 5.5, it's a waitlist research preview for US Pro and Max users.
Official Claude Duet app icon and wordmark from Anthropic, above the headline "Duet is a proactive Claude, on a waitlist."

ALT Official Claude Duet app icon and wordmark from Anthropic, above the headline "Duet is a proactive Claude, on a waitlist."

1
44
Anthropic announced Claude Duet on October 21. Instead of waiting for prompts, Claude looks for what needs your attention and checks with you before it acts. It runs on Opus 5.5 and carries over your existing Claude memory, skills and connected apps.
Text poster: "It looks for work instead of waiting for prompts." Memory, skills and connected apps carry over.

ALT Text poster: "It looks for work instead of waiting for prompts." Memory, skills and connected apps carry over.

1
22
Catch up: @n8n_io released Nod yesterday, a free, open-source personal agent with its own sandboxed computer and browser. It runs on your Mac or PC, works in Gmail, Slack and sites with no API, and asks before it sends, pays or deletes. The privacy docs have a catch.
n8n Labs launch art reading "Introducing Nod, a standalone personal agent from n8n Labs that runs on your computer" above the headline "n8n built an agent with its own computer."

ALT n8n Labs launch art reading "Introducing Nod, a standalone personal agent from n8n Labs that runs on your computer" above the headline "n8n built an agent with its own computer."

1
20
The fine print: usage data is on by default. Screens and features used, task types, timings, costs, failures and onboarding answers go to n8n's PostHog project in the EU. Your messages to Nod and its replies are excluded. Conversation traces are off by default.
Nod's in-app "How Nod uses your data" panel showing Usage data On and Conversation traces Off, under the headline "Usage data ships on by default."

ALT Nod's in-app "How Nod uses your data" panel showing Usage data On and Conversation traces Off, under the headline "Usage data ships on by default."

1
17
Catch up: @ElevenLabs shipped Scribe v3, its new speech-to-text model. It makes 27% fewer word errors than Scribe v2 across 102 languages: 12.4% vs 17.0%. API price per hour of audio: Scribe v3 $0.20, Scribe v2 $0.20. The catch is in how that 27% was measured.
ElevenLabs' official Scribe v3 cover art: the words Scribe v3 in white over a blue and white blurred gradient. Headline: ElevenLabs ships Scribe v3.

ALT ElevenLabs' official Scribe v3 cover art: the words Scribe v3 in white over a blue and white blurred gradient. Headline: ElevenLabs ships Scribe v3.

1
22
Word error rate across all 102 languages, lower is better: Scribe v3 12.4%, Scribe v2 17.0%, Gemini 3.5 Transcribe 18.5%. Gains were bigger in harder languages: Swahili errors fell 34%, Tamil 29%, Thai 28%. English improved 23%.
Bar chart of word error rate across 102 FLEURS languages, lower is better: Scribe v3 12.4%, Scribe v2 17.0%, Gemini 3.5 Transcribe 18.5%. Headline: The lowest error rate.

ALT Bar chart of word error rate across 102 FLEURS languages, lower is better: Scribe v3 12.4%, Scribe v2 17.0%, Gemini 3.5 Transcribe 18.5%. Headline: The lowest error rate.

1
17
Just in: @GoogleDeepMind released AlphaGenome 2, a DNA model that predicts what a single-letter mutation does inside one specific cell type instead of a blurry tissue average. It trained on 107M+ single-cell profiles across 68 tissues. The catch sits in the API terms.
Google DeepMind's official AlphaGenome 2 launch art: glowing pink cubes subdivided into smaller cubes on a dark purple background. Headline: DNA, read cell by cell.

ALT Google DeepMind's official AlphaGenome 2 launch art: glowing pink cubes subdivided into smaller cubes on a dark purple background. Headline: DNA, read cell by cell.

8
Just in: @GoogleDeepMind released AlphaGenome 2, a DNA model that predicts what a single-letter mutation does inside one specific cell type instead of a blurry tissue average. It trained on 107M+ single-cell profiles across 68 tissues. The catch sits in the API terms.
Google DeepMind's official AlphaGenome 2 launch art: glowing pink cubes subdivided into smaller cubes on a dark purple background. Headline: DNA, read cell by cell.

ALT Google DeepMind's official AlphaGenome 2 launch art: glowing pink cubes subdivided into smaller cubes on a dark purple background. Headline: DNA, read cell by cell.

1
10
On TenK10K, a single-cell map of 28 immune cell types, it found 30% of causal variants vs 16% for the best prior tool. On the direction of each gene effect: 43% vs 23%. Researchers can also condition it on a few dozen cells from their own experiment, no retraining.
Bar chart on TenK10K causal variants found: AlphaGenome 2 30%, best prior tool 16%. Headline: Nearly twice the hits.

ALT Bar chart on TenK10K causal variants found: AlphaGenome 2 30%, best prior tool 16%. Headline: Nearly twice the hits.

1
6
Catch up: @SpaceX's solar deal with SINGULUS has a twist in the fine print. The German supplier will ship a turnkey heterojunction solar cell line to Bastrop, Texas, in H1 2027, for Starmind satellites. That one order is about as big as SINGULUS's sales for nine months.
SINGULUS GENERIS PECVD solar cell coating line in a factory hall, under the headline "SpaceX orders a solar cell line."

ALT SINGULUS GENERIS PECVD solar cell coating line in a factory hall, under the headline "SpaceX orders a solar cell line."

1
25
The order: over €60M for the first line. SINGULUS booked €60.4M of revenue from January to September and expects about €83M for all of 2026. Its whole order backlog on Sept 30 was €47.1M. This one contract is bigger.
Bar chart in millions of euros: SpaceX first line over 60, SINGULUS January to September revenue 60.4, order backlog on Sept 30 47.1, under the headline "One order is worth nine months of sales."

ALT Bar chart in millions of euros: SpaceX first line over 60, SINGULUS January to September revenue 60.4, order backlog on Sept 30 47.1, under the headline "One order is worth nine months of sales."

1
16
Catch up: @cursor_ai rebuilt Projects for multiplayer. Your coding agents can now coordinate with your teammates' agents in one workspace. 435K views on the launch post. The blog says beta for all users. The docs add the conditions.
Cursor Projects task board for a "Billing refactor" project with tasks grouped as Ready to Merge, For Review and Working, each linked to a PR and a teammate avatar, under the headline "Your agents now talk to your teammates' agents."

ALT Cursor Projects task board for a "Billing refactor" project with tasks grouped as Ready to Merge, For Review and Working, each linked to a PR and a teammate avatar, under the headline "Your agents now talk to your teammates' agents."

1
6
How it works: each teammate keeps a private chat with their own coordinator and worker agents. On top sits a shared task board (who is doing what, and the PR behind each task) plus shared context like specs and mockups. A decision in one chat reaches every agent in the Project.
Cursor's "Invite teammates" panel listing team members, under the headline "Sharing needs a team account."

ALT Cursor's "Invite teammates" panel listing team members, under the headline "Sharing needs a team account."

1
7
Catch up: @cursor_ai rebuilt Projects for multiplayer. Your coding agents can now coordinate with your teammates' agents in one workspace. 435K views on the launch post. The blog says beta for all users. The docs add the conditions.
Cursor Projects task board for a "Billing refactor" project with tasks grouped as Ready to Merge, For Review and Working, each linked to a PR and a teammate avatar, under the headline "Your agents now talk to your teammates' agents."

ALT Cursor Projects task board for a "Billing refactor" project with tasks grouped as Ready to Merge, For Review and Working, each linked to a PR and a teammate avatar, under the headline "Your agents now talk to your teammates' agents."

4
Catch up: @stripe closed its OpenRouter acquisition yesterday, 83 days after agreeing to it. @OpenRouter is the gateway 10M+ developers use to reach 400+ AI models through one API. The price was never disclosed. The NYT put it at about $7.5B.
OpenRouter and Stripe logos side by side above the headline "Stripe now owns OpenRouter."

ALT OpenRouter and Stripe logos side by side above the headline "Stripe now owns OpenRouter."

1
16
OpenRouter's reported value: about $0.55B at its June 2025 Series A (PitchBook), $1.3B at its May 2026 Series B (NYT), about $7.5B in the Stripe deal (NYT). Roughly 14x in 14 months. Stripe declined to comment on terms.
Bar chart of OpenRouter's reported valuation: $0.55B Series A June 2025, $1.3B Series B May 2026, $7.5B Stripe deal August 2026.

ALT Bar chart of OpenRouter's reported valuation: $0.55B Series A June 2025, $1.3B Series B May 2026, $7.5B Stripe deal August 2026.

1
16
Catch up: Dutch lender ING is putting about $500M of data-center loans up for sale via BWIC, including a $25M slice tied to the New Mexico Stargate campus for @OpenAI, @SoftBank, and @Oracle.
ING House in Amsterdam's Zuidas fading into espresso; headline ING is selling $500M in data-center loans.

ALT ING House in Amsterdam's Zuidas fading into espresso; headline ING is selling $500M in data-center loans.

1
9
How it works: about a dozen loans in a bids-wanted-in-competition process. Minimum bid size $10M. Offers due Nov 19. ING declined to comment when asked by CNBC and Reuters.
Slide on the $25M New Mexico Stargate loan slice.

ALT Slide on the $25M New Mexico Stargate loan slice.

1
6
Load more