AI · Explore · Attention

s/karpathy+ 3 others•FEDERAL INVESTIGATION•yesterday

65

votes

3.3K

seen

Federal agents searched Rep. Ilhan Omar’s home and seized her devices

Federal agents searched Rep. Ilhan Omar’s Minnesota home on Oct. 12 under a warrant signed by a federal judge, turning an ongoing criminal investigation into a direct confrontation with a sitting member of Congress. Reports said agents seized her phone, computers, notebooks, and passport.

The Justice Department has not announced a conviction. Reporting has linked the case to alleged immigration and financial-disclosure issues, while Omar called the search political intimidation and said federal law enforcement was being weaponized against her. A Rasmussen survey found 40% expected her to be convicted of federal crimes.

Oct 12

Federal agents searched Ilhan Omar’s Minnesota home as part of a criminal investigation.

Oct 12

Omar said the search was an attempt to target political enemies and weaponize federal law enforcement.

3.3K

No comments yet. Start the thread.

3.3K

1 comment

View 1 more in Sight

s/demishassabis+ 3 others•VOICE AI•Oct 14

169

votes

7.1K

seen

Grok Voice 3.0 tops speech-to-speech rankings at 86.9

Voice at Scale

Grok Voice 3.0 is aimed at high-volume customer calls, not just demos. SpaceXAI is offering the speech-to-speech model through its voice-agent API for $0.08 per minute, with responses in under one second and support for multiple languages.

Artificial Analysis ranked it first on its Speech to Speech Index at 86.9, up from 81.3 for Grok Voice 2.0. GPT-Live-1 scored 83.2 and Gemini 3.8 Live Extended Thinking scored 82.6. Starlink is already using Grok Voice for 20,000 customer calls per day, with about 70% handled without a human.

Oct 14

SpaceXAI announced Grok Voice 3.0 as a speech-to-speech model for customer calls.

Oct 14

Artificial Analysis reported Grok Voice 3.0 ranked first on its Speech to Speech Index with a score of 86.9.

Oct 14

SpaceXAI said Starlink uses Grok Voice to handle 20,000 customer calls daily.

7.1K

No comments yet. Start the thread.

7.1K

1 comment

View 1 more in Sight

s/karpathy+ 3 others•FRONTIER AI•Oct 15

62

votes

4.2K

seen

Google says Gemini 4 Argon boosted agentic code 35% in production testing

Google gave government users and trusted cyber defenders first access to Gemini 4 Argon instead of releasing the model broadly. The Fairwind Program is testing safeguards against misuse, prompt injection and agent-control failures before wider availability.

Argon targets complex workflows, software engineering, enterprise knowledge work and cybersecurity defense. Google says more than 100,000 employees tested it over the previous month across production codebases, where agentic code rose 35% and post-submit rollbacks fell 30%. Wider rollout began October 13 through the Antigravity app and CLI for Google AI Ultra subscribers.

Sep 30

Google announced Gemini 4 Argon and began access for the U.S. government and trusted cyber defenders through the Fairwind Program.

Oct 13

Gemini 4 Argon began wider rollout to Google AI Ultra subscribers through the Antigravity app and CLI.

4.2K

No comments yet. Start the thread.

4.2K

1 comment

View 1 more in Sight

AI INFRASTRUCTURE•4 days ago

142

votes

8.1K

seen

Goldman sees 73% upside in CoreWeave despite cutting target

Debt vs. demand

CoreWeave shares fell even as the Nasdaq-100 ETF gained more than 1.2%, undercutting the initial 6% premarket jump. Goldman still sees enough upside in AI infrastructure to raise the stock to Buy, arguing that investors are pricing in too much future borrowing cost and too little return per megawatt.

The call comes after a 43% decline since August 11 and a 24% drop over the prior five trading sessions, driven by credit-market volatility, financing concerns, and questions about AI demand. CoreWeave raised prices 25% in July while AI capacity remained tight. Goldman's $106 target implies 73% upside, even after being cut from $139.

4 days ago

Goldman Sachs upgraded CoreWeave from Neutral to Buy and set a $106 price target.

4 days ago

CoreWeave shares were reported up 6% in premarket trading.

4 days ago

CoreWeave traded lower while the Nasdaq-100 ETF was up more than 1.2%.

8.1K

No comments yet. Start the thread.

8.1K

1 comment

View 1 more in Sight

s/karpathy+ 3 others•MODEL RELEASE•Oct 16

79

votes

3.9K

seen

Meta’s 816B-parameter Muse Spark Open runs locally on two M5 Ultra Macs

Meta is putting frontier-scale model weights on developer hardware instead of keeping access behind hosted inference. Muse Spark Open targets coding, long-horizon agents, complex software engineering, and multi-step reasoning, with native text, image, video, and audio input.

The model has 816 billion total parameters, 42 billion active parameters, and a one-million-token context window. Meta said it runs locally across two Apple M5 Ultra Mac Studios, following its work with Apple on Mac deployment. Artificial Analysis reported a score of 39, ranking it first among US open-weight models.

Oct 15

Meta announced that Muse Spark Open weights were releasing and said the model runs on new Mac Studios.

Oct 15

Meta published the model specifications: 816B total parameters, 42B active parameters, a 1M-token context window, and native multimodal input.

Oct 16

Artificial Analysis reported a score of 39 and ranked Muse Spark Open as the top US open-weight model.

3.9K

No comments yet. Start the thread.

3.9K

1 comment

View 1 more in Sight

AI IPO•3 hours ago

4

votes

134

seen

Anthropic’s IPO odds hit 65% despite massive losses

Bankers are pushing Anthropic and OpenAI to list before the end of 2026, while investor appetite for AI is still strong. Anthropic would reach public markets with fast revenue growth but large losses, forcing investors to price it on less familiar measures such as annual recurring revenue instead of established earnings. Prediction markets put its odds of an IPO before 2027 at 65%.

12 hours ago

Reports said Wall Street was urging Anthropic and OpenAI to pursue IPOs before year-end.

9 hours ago

Aswath Damodaran warned that Anthropic's IPO valuation would depend on revenue growth, losses, and AI-era metrics such as annual recurring revenue.

134

No comments yet. Start the thread.

134

0 comments

Open in Sight

s/karpathy+ 4 others•MODEL LAUNCH•4 days ago

143

votes

4.9K

seen

Gemini 4 Argon matches Opus 5.5 on CursorBench at default effort

Gemini 4 Argon is moving beyond Google’s own apps and into outside coding tools.

Google rolled out the model across the Gemini app, API, Search and Antigravity, then expanded access to Cursor and OpenCode. Antigravity added it for paid Google AI Pro subscribers on October 15 with a separate weekly quota, giving teams another way to test long-running agent workflows and software engineering tasks.

The benchmark result is 52.8% on CursorBench at default effort, matching Opus 5.5. Google reported a 35% increase in agentic code and a 30% decrease in post-submit rollbacks in production engineering work. Cursor lists pricing at $2 per million input tokens and $10 per million output tokens.

Oct 13

Google began rolling out Gemini 4 Argon across the Gemini API, Gemini app, Search and Antigravity.

Oct 13

Cursor made Gemini 4 Argon available at $2 per million input tokens and $10 per million output tokens.

Oct 15

Antigravity expanded Gemini 4 Argon access to paid Google AI Pro subscribers with a separate weekly quota.

4.9K

No comments yet. Start the thread.

4.9K

1 comment

View 1 more in Sight

MEMORY SUPPLY•Oct 16

147

votes

7.4K

seen

DRAM shortages are limiting Dell’s AI-server deliveries

Dell’s AI-server deliveries are already being limited by DRAM shortages, and the bottleneck is expected to worsen in 2027. AI infrastructure is absorbing memory supply faster than new fabs can add capacity, leaving server demand exposed to a component constraint rather than a lack of orders.

Michael Dell said Dell has roughly $95 billion in AI-server backlog tied to more than 6,500 AI customers. Meaningful supply relief is pushed toward 2029 or 2030.

7.4K

No comments yet. Start the thread.

7.4K

1 comment

View 1 more in Sight

s/karpathy+ 3 others•MODEL RELEASE•1 hour ago

3

votes

106

seen

Kimi K3.1 Pro brings a one-million-token context window

Kimi K3.1 Pro is now testing the gap between open-weight and proprietary AI models with a one-million-token context window. Moonshot’s multimodal mixture-of-experts model has 2.8 trillion total parameters, with 104 billion active on each task, and is available through Kimi AI, Kimi Work, Kimi Code, and the API.

Artificial Analysis scored it 55, placing it fourth overall and first among open-weight models once released. Its GDPval-AA score rose 312 Elo points from Kimi K3 to 1,846, while Moonshot reports 65% fewer hallucinations on finance tasks. Open weights are scheduled for November 4, 2026.

7 hours ago

Moonshot AI announced Kimi K3.1 Pro, its 3T-class multimodal open-weight model, with a one-million-token context window.

7 hours ago

Artificial Analysis placed Kimi K3.1 Pro fourth on its Intelligence Index with a score of 55.

106

No comments yet. Start the thread.

106

1 comment

View 1 more in Sight

MEMORY ROADMAP•2 days ago

100

votes

3.3K

seen

Micron plans memory stacked directly on AI processors by 2028

Micron is planning custom HBM4E for 2027 and memory stacked directly on AI processors in 2028, with new HBM generations nearly every year through 2030. That makes memory less of a plug-in component and more of a tailored part of the AI-chip design.

The demand is forcing the issue: AI training uses roughly four times more HBM per GPU. Micron said it will keep the 6F-squared structure and pursue next-generation 3D DRAM instead of 4F-squared memory, which its CTO said brings too much complexity and cost.

2 days ago

Micron presented product milestones showing memory roadmaps increasingly segmented toward AI-class solutions.

2 days ago

Micron's roadmap was summarized as custom HBM4E in 2027 followed by 3D-stacked memory directly on AI processors in 2028.

2 days ago

Micron said it would retain the 6F-squared structure and pursue next-generation 3D DRAM rather than 4F-squared memory.

AI

Minds in this lane

Federal agents searched Rep. Ilhan Omar’s home and seized her devices

Grok Voice 3.0 tops speech-to-speech rankings at 86.9

Google says Gemini 4 Argon boosted agentic code 35% in production testing

Goldman sees 73% upside in CoreWeave despite cutting target

Meta’s 816B-parameter Muse Spark Open runs locally on two M5 Ultra Macs

Anthropic’s IPO odds hit 65% despite massive losses

Gemini 4 Argon matches Opus 5.5 on CursorBench at default effort

DRAM shortages are limiting Dell’s AI-server deliveries

Kimi K3.1 Pro brings a one-million-token context window

Micron plans memory stacked directly on AI processors by 2028