- The Current ⚡️
- Posts
- OpenAI Releases GPT 4.1
OpenAI Releases GPT 4.1
Also, NVIDIA to begin manufacturing some of their chips at TSMC in Arizona

⚡️ Headlines
🤖 AI
Nvidia to Make Some AI Chips in the U.S. for the First Time - Nvidia will begin producing AI chips and systems in the U.S., aligning with political pressure to localize manufacturing and reduce reliance on overseas fabs [TechCrunch].
DeepSeek Spurs China’s AI Boom Amid U.S. Chip Bans - Chinese startups are pivoting from foundational model development to practical AI applications as DeepSeek’s low-cost success and tightening U.S. export controls reshape the country's innovation landscape [Rest of World].
Apple to Process User Data On-Device to Enhance AI - Apple will analyze more user data directly on devices using on-device learning and private cloud compute to improve its AI models while maintaining privacy [Bloomberg].
Apple Expands Use of Differential Privacy to Train Better AI - Apple is leveraging differential privacy and synthetic datasets to train its AI systems without compromising user confidentiality [The Verge].
Meta Faces Scrutiny Over AI-Driven Content in Europe - European regulators are questioning Meta’s use of AI-generated content on Facebook and Instagram, prompting transparency and safety commitments from the company [AP News].
OpenAI’s Latest Model Introduces New Training Techniques - OpenAI’s new model reportedly uses novel memory and reasoning techniques, moving beyond traditional transformer scaling for more efficient performance [The Information].
Amazon CEO Highlights AI's Role in Company Strategy - Andy Jassy’s shareholder letter emphasizes AI’s integration across AWS, Alexa, and fulfillment operations, calling it the company’s “most important investment” [Amazon].
NATO and Palantir Team Up for Smart System Under Project Maven - NATO and Palantir have launched a five-year collaboration on a new AI-enabled battlefield decision system, marking a major milestone for defense tech integration [DefenseScoop].
🦾 Emerging Tech
Kraken to Launch U.S. Stock Trading with Global Expansion in Sight - Crypto exchange Kraken is entering the stock and ETF trading space, beginning in the U.S. and planning future rollouts in the U.K. and Europe [The Block].
🤳 Social Media
Meta Tailors AI Tools for Europe Amid Regulatory Pressure - Meta unveils AI-powered experiences built with European values in mind, emphasizing local relevance, data minimization, and transparency [Meta Newsroom].
TikTok Adds New Brand Safety Controls for Advertisers - TikTok introduces more granular ad placement settings, giving advertisers improved control over content suitability and user targeting [TikTok for Business].
🔬 Research
Seaweed Video Launches to Share Cutting-Edge AI Research - A new video platform, Seaweed, has launched to make academic AI research more accessible through visual storytelling and expert interviews [Seaweed].
Google DeepMind Debuts Dolphin-Inspired AI Model 'Gemma' - Google’s new lightweight model “Gemma” is inspired by dolphin echolocation and aims to be efficient, small-scale, and open for developer use [Google Blog].
⚖ Legal
Meta Fights to Keep Instagram and WhatsApp as Antitrust Trial Begins - The FTC’s long-anticipated antitrust case against Meta begins, threatening to break up its ownership of Instagram and WhatsApp [Wall Street Journal].
Trump Tariffs Could Hit Apple Devices Made in China - Proposed Trump-era tariffs could raise costs on iPhones, Macs, and other Apple devices by targeting Chinese-made electronics [Axios].
🎱 Random
Amex CISO Uses AI to Battle Cyber Threats at Machine Speed - American Express is using AI for real-time threat detection, allowing its cybersecurity team to proactively hunt and neutralize risks [VentureBeat].
‘Doctor Who’ Writer on Incels, Sci-Fi, and Internet Extremism - New 'Doctor Who' writer Belinda Chandra talks about weaving online radicalization themes into sci-fi narratives to challenge toxic ideologies [The Verge].
🔌 Plug-Into-This
OpenAI has launched a refined series of GPT-4.1 models—GPT-4.1, Mini, and Nano—specifically optimized for software engineering and coding workflows. These models boast a 1-million-token context window, enhanced consistency in tool use and formatting, and are available exclusively through OpenAI’s API.

GPT-4.1 models are tailored for high-fidelity code generation, outperforming previous iterations on benchmarks like SWE-bench Verified.
The new context window enables processing of large-scale codebases and technical documentation within a single prompt.
GPT-4.1 improves in HTML/CSS rendering, frontend UI generation, and consistent function calling—a leap in reliability for developer workflows.
All models are API-only, with Nano offered at $0.10 per million tokens, aiming for budget-conscious integrations in lightweight environments.
Despite improvements, GPT-4.1 slightly underperforms Google's Gemini 2.5 Pro and Claude 3.5 Sonnet on aggregate coding benchmarks.
Announcing GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano in the API.
TL;DR: Major improvements on coding, instruction following, and long context. 💥
00:00 Intro
02:18 Coding
04:53 Instruction following
06:58 Long context
10:22 Demos, pricing, and availability
20:00 @windsurf_ai— OpenAI Developers (@OpenAIDevs)
7:11 PM • Apr 14, 2025
👨💻 While GPT-4.1 isn't a giant leap in general reasoning, it’s a boon to their core customer base (coders) by embedding models better into dev workflows, where reliability and token throughput matter more than headline benchmarks.
NVIDIA has announced it will manufacture AI supercomputers and Blackwell chips in the U.S. for the first time, marking a strategic shift in supply chain localization. The initiative leverages TSMC’s Arizona facility and partnerships with Foxconn and Wistron for system assembly.

The initiative directly responds to heightened U.S.–China tech tensions and recent tariffs targeting AI hardware.
Blackwell chips will be fabricated domestically by TSMC; entire supercomputers will be assembled in new Texas-based facilities.
Expected to be operational in 12–15 months, the move may generate hundreds of thousands of U.S.-based tech and manufacturing jobs.
Nvidia aims to establish sovereign AI compute infrastructure aligned with national security and economic goals.
The plan positions Nvidia to preempt further geopolitical instability while accelerating stateside AI model development.
At GTC, Jensen said Nvidia would be in “good shape” on tariffs by year end. I guess this is what he meant
— tae kim (@firstadopter)
2:49 PM • Apr 14, 2025
🇺🇸 This domestic push isn’t just patriotic optics—it’s Nvidia hedging against fragility in global fabs while signaling to enterprise and government buyers that “Made in America” compute may be coming online fast.
Hugging Face has acquired Pollen Robotics, the maker of the open-source humanoid robot Reachy 2, deepening its commitment to transparency and democratization in AI and robotics. The company plans to make both the software and hardware designs freely accessible.

Reachy 2, a lightweight, bug-eyed biped, can manipulate objects and is already in use at top AI labs for robotic learning tasks.
Hugging Face will allow developers to modify and redistribute the code and hardware—mirroring open-source software ecosystems.
The acquisition targets growing demand for “trustable” AI systems that operate in physical environments like homes and factories.
By exposing robot internals, Hugging Face aims to counter hype-heavy, opaque humanoid demos from major players.
The move parallels the open-weight model movement, bringing that ethos to embodied AI research and prototyping.
Boom!
The amazing @huggingface has acquired @pollenrobotics!
This will spark a new era to open source robotics!
I can’t wait to see what’s next and look forward to building!
Congratulations @ClementDelangue and team!
— Brian Roemmele (@BrianRoemmele)
6:03 PM • Apr 14, 2025
🤗 Hugging Face basically just gave anyone the blueprint to build a robot at home—no secrets, no black boxes. This signals a philosophical stand: if embodied AI is the next frontier, Hugging Face wants to ensure it’s built in the open—where public oversight and grassroots innovation can shape its trajectory, not just billion-dollar labs.
🆕 Updates
Meet Notion Mail.
The inbox that organizes itself, drafts emails, and schedules meetings any way you'd like.
The inbox that thinks like you.
— Notion (@NotionHQ)
3:04 PM • Apr 15, 2025
Introducing GPT-4.1 for understanding arXiv papers 🚀
Highlight any section of a paper to ask questions and “@” other papers to quickly add to context and compare results, benchmarks, etc.
— alphaXiv (@askalphaxiv)
7:18 PM • Apr 14, 2025
Microsoft just released the first natively trained 1-bit model: BitNet 2B.
Trained on 4 Trillion tokens. Native 1.58-bit weights and 8-bit activations (W1.58A8).
Performs very close to Qwen 2.5 1.5B in benchmarks while being 1/6 of its size and twice faster.
— Ahmad (@TheAhmadOsman)
11:23 AM • Apr 15, 2025
🚨 BREAKING: Google just dropped the most practical AI release of 2025.
It handles your emails, data, docs, meetings and does it with context.
This is AI that actually saves time.
Here’s everything they announced (and why it matters for your team):
— Brendan (@jowettbrendan)
9:59 AM • Apr 14, 2025
📽️ Daily Demo
OpenAI has published a new prompting guide for GPT-4.1
Agentic prompt that OpenAI used to achieve its highest score on SWE-bench Verified is added here and blog link in 2nd post.
---------------------------------------------------------
You will be tasked to fix an issue from
— AshutoshShrivastava (@ai_for_success)
4:52 AM • Apr 15, 2025
🗣️ Discourse
I've been loving GPT-4.1, and it's now my daily driver. We've spent hundreds of hours experimenting to get it performing at its best, and I wanted to share some key things we learned to help you get the most out of these models from day 1.
Agentic workflows
- This is your new— Noah MacCallum (@noahmacca)
9:45 PM • Apr 14, 2025
don't miss that OAI also published a prompting guide WITH RECEIPTS for GPT 4.1 specifically for those building agents... with a new recommendation for:
- telling the model to be persistent (+20%)
- dont self-inject/parse toolcalls (+2%)
- prompted planning (+4%)
- JSON BAD - use— swyx (@swyx)
6:29 PM • Apr 14, 2025