| |
Welcome, humans. | We never thought we would learn so much about how to think about Taiwan endgame scenarios on a cooking show, but thanks to Latent Space's new In-Context Cooking and SemiAnalysis founder (codename: whale watcher) Dylan Patel, we sure did! | | If you're wondering how China's interest in Taiwan and TSMC (the company who makes the most sophisticated chips for NVIDIA, Apple, and others) could play out, and what the best scenario would be for maintaining "the status quo", plus how to think about AI CAPEX (the hyperscalers know their levels of demand better than you do, but don't necessarily bet against the market "getting really mad" at them), check this one out. | Here's what happened in AI today: | The AI landscape is moving on every front at once—agents, benchmarks, chips, and open source. Nvidia is unveiling a new inference chip with Groq's technology at GTC, with OpenAI as a major customer. Google struck a multibillion-dollar deal to supply Meta with TPU AI chips, challenging Nvidia. SemiAnalysis argues Anthropic will grow revenue faster than OpenAI from here on out.
| | P.S: Yesterday we missed the links for Anthropic's 32 page comple Skills explainer and our write-up on it; they're on the web version now, but sorry we missed it in the email!! | | The AI Landscape Is Chaos. Here's the One Recap You Actually Need. | FULL NEWS BRIEF: Here's our attempt to explain everything going on in AI right now | AI is simultaneously getting smarter, cheaper, faster, and more useful all at once, across every front. And the companies building it are making very different bets on who they're building it for. | The big splits happening right now: | Developers vs. everyone else. OpenAI is laser-focused on coders with 1M+ developers are now using Codex, their cloud-based software engineering agent. Meanwhile, Anthropic is going hard after office workers and enterprise, with 80% of revenue now B2B and 13 new enterprise plugins (Google Workspace, DocuSign, FactSet) via Claude Cowork. But devs are pushing back: Anthropic recently blocked popular third-party tools like OpenCode and Cursor, prompting some to publicly switch to Codex. Benchmarks are already broken. ARC-AGI-2 was specifically designed to resist AI brute-forcing. When it launched, pure LLMs scored 0%. Today, GPT-5.2 sits at ~53%, Gemini 3 Deep Think hit 84.6%, and one method using Gemini pushed to 95.1%. ARC-AGI-3, which shifts to interactive and video-game-like environments, isn't officially out until March 25. People are already cracking the preview. Agents are everywhere. Every major AI platform now has an agent layer. Copilot Studio has 1,400+ connectors. Gemini's Opal 2.0 makes decisions autonomously. Claude has Cowork. And Perplexity just launched Computer, a $200/month super-agent that juggles 19 different AI models at once. Big tech is betting massive on infrastructure. Companies are set to spend $650B-$690B on data centers in 2026 alone, a 67-74% increase year over year. But the more interesting story is efficiency. New chip startups like MatX (raised $500M, targeting 2,000 tokens/second) and Cerebras (confidentially filed for IPO at a $23B valuation) are trying to make the chips we already have work dramatically harder.
| The open-source agent story alone is worth the click. One project called OpenClaw hit 100,000+ GitHub stars in days. A user's agent negotiated $4,200 off a car purchase while he slept. Another accidentally won an insurance dispute. Chinese companies cloned and productized it almost overnight, and security researchers are already sounding alarms, calling the data exposure risk deeper than TikTok. | And that's before we even get into diffusion language models running 5-10x faster than traditional LLMs, post-transformer architectures, or why Perplexity just abandoned its entire ad business to bet everything on subscriptions. | |
|
|
|
|
Watch this demo to see how Slackbot: | Makes your entire workspace searchable (docs, convos, apps) Enhances every teammate with role-specific automations Learns your project and preferences over time for even smarter outputs
| No onboarding or setup. Just start chatting with Slackbot like another teammate. | Watch this 2-minute demo. | |
Prompt Tip of the Day | Greg Isenberg stress-tested Perplexity Computer live and turned it into a one-person sales, research, and strategy team. The highlights: | Cold outreach: It found real email addresses, researched each prospect's recent activity, wrote personalized pitches, and sent them via connected Gmail—all in parallel. Competitive intel: A recurring 8 AM monitor across five competitors. New episodes, pricing changes, social buzz. Nothing changed? No notification. VC research: Described his startup, told it to find 50 best-fit investors. Got a spreadsheet with fund sizes, partner contacts, and thesis alignment. Investment memos: Asked for a full Shopify analysis with charts, comparisons, and bull / bear cases. Got a PDF.
| Our favorite insight: When Greg asked, "Is there something I'm not asking you that could make me more money?"—it suggested tracking competitors' new sponsors so he could reach out while their podcast budget was still hot. Try that prompt in your next AI session. | | Treats to Try | *Asterisk = from our partners (only the first one!). Advertise to 650K+ readers here! | |
|
| Around the Horn |  | This is awesome. |
| Nvidia plans to unveil a new inference-focused processor incorporating Groq's chip technology at next month's GTC conference, with OpenAI as a major customer following a $20B licensing deal. Google struck a multibillion-dollar deal to supply Meta with its TPU AI chips, directly challenging Nvidia's dominance in the accelerator market. SemiAnalysis' Doug O'laughlin and team wrote a deep dive on why Claude Code is so good, breaking down Claude Code's role as an inception point for agentic systems, and extrapolate that Anthropic will grow revenue faster than OpenAI from here on out. Reiner Pope of MatX, the best new AI chip startup, went on the Cheeky Pint podcast to talk all about AI chips; if you've ever wondered what a GPU is, how it's different from CPUs, and what the future of AI chips could (very likely) look like, you'll want to watch this one! Corey's model routing guide breaks down how to slash your AI token bill 40-60% by sorting every task into three tiers (premium, workhorse, utility) instead of running one expensive model for everything. Watch: One more Latent Space podcast plug for ya: Joel Becker of METR Research breaking down the most infamous chart in AI benchmarking; the long-horizon task completion chart; make sure to skim the write-up that goes along with this, too.
| |
|
| Coresignal's Data Search enables non-technical professionals to create high-quality B2B lists and datasets in a few simple steps: | Type your search in plain English Get a preview of up to 100 matched records Enrich selected or all records instantly Generate a dataset of up to 10,000 records in JSON
| The result is a clean, structured company or employee list that is ready for CRM import, outreach, analytics, AI platform development, or internal tools. | Try out Data Search for free | |
|
 | AGI can't be worse at humanity than we are. |
| | |
| | | That's all for now. | | What'd you think of today's email? | |
|
| P.S: Before you go… have you subscribed to our YouTube Channel? If not, can you? |  | Click the image to subscribe! |
| P.P.S: Love the newsletter, but only want to get it once per week? Don't unsubscribe—update your preferences here. |
|
No comments:
Post a Comment
Keep a civil tongue.