All Updates
Page 748 of 1667
May 4, 2026
SSMs Struggle in 25M Param Training
Empirical findings from OpenAI's Parameter Golf reveal SSMs compress worse than transformers and lose architectural edges at larger vocabs in tiny models (25M params, 16MB, 10min train). Blog details Mamba-3 Triton kernel experiments uncovering fusion slowdowns, quant bugs, and precision fixes.
YouTube Auto-Dubbing Lacks Off Switch
YouTube has introduced auto-dubbing that automatically translates and dubs videos into other languages like Japanese without user consent. Viewers experience unwanted dubs, such as hearing Marques Brownlee speak Japanese. Crucially, there is no global off switch to disable this feature.
New Mexico Court: Meta Teen Harm Trial Begins
New Mexico lawsuit against Meta opened in court this week. Accuses Facebook, Instagram, WhatsApp of harming teen mental health. State seeks injunction forcing platform reforms.
Palantir Earnings Test in AI Sell-Off
Palantir's Q1 earnings arrive after a 30% YTD stock drop amid AI software sell-off. It's the first chance to prove resilience against multiple compression affecting peers. Until recently, it led large-cap AI software performers over Salesforce, Microsoft.
NeurIPS 2025 Creative AI Papers Missing
NeurIPS 2025 announced the Creative AI Track as part of official proceedings with poster presentations. However, the live proceedings lack these papers. Community seeks explanation for the discrepancy.
Cerebras Cuts IPO to $3.5B Raise at $26.6B Valuation
Cerebras Systems, an AI chipmaker, has filed updated IPO terms to raise $3.5 billion at a $26.6 billion valuation. This is lower than the initially suggested $4 billion at $40 billion, aligning instead with its February private valuation. The offering includes 28 million shares priced at $115-$125 each.
Goldman, Blackstone Partner with Anthropic on AI Firm
Anthropic PBC is partnering with Blackstone Inc., Hellman & Friedman, and Goldman Sachs Group Inc. to form a new company. The firm will deliver artificial intelligence services to a broad group of midsize companies.
Instagram Tests Optional AI Creator Labels
Instagram is testing optional 'AI creator' labels for accounts frequently posting generative AI content. The company encourages use of the feature but does not require it. This helps users identify AI-generated posts more easily.
DoorDash Adds AI for Merchant Onboarding & Photos
DoorDash has introduced AI-powered tools to streamline merchant operations. Merchants can now speed up onboarding, edit dish photos for better appeal, and generate websites from existing content. This aims to reduce friction for restaurant partners joining the platform.
Llama.cpp MTP Support Hits Beta
Llama.cpp now offers beta support for MTP, starting with Qwen3.5, thanks to contributor Aman. This feature, combined with maturing tensor-parallel support, aims to close performance gaps with vLLM in token generation speeds. Merge into mainline is expected soon.
AI Slop Pollutes Podcasts: 39% Machine-Made
A wave of AI-generated podcasts is surging through audio platforms. Bloomberg reports nearly 39% of new podcast feeds over nine days were machine-made. This shows AI 'slop' extending beyond music to podcasts.
Google CEO Meets White House on TPU Approvals
Google CEO Sundar Pichai visited the White House for high-level meetings with Trump administration officials. The core issue is US concerns over insufficient classified AI compute capacity for national defense. The government plans to accelerate TPU security approval processes.
AMD Ryzen AI MAX 495 Leaks with iGPU Upgrade
AMD is preparing the Ryzen AI MAX 400 series, codenamed Gorgon Halo, as a refresh of Strix Halo. It retains Zen 5 CPU and RDNA 3.5 GPU on FP11 platform with higher clockspeeds and larger memory support.
Automating AI Research: Self-Improvement Step
Import AI newsletter issue 455 covers automating AI research. It frames this as the first step towards recursive self-improvement in AI systems. The edition highlights progress in making AI R&D more autonomous.
OpenAI Launches $10B AI Deployment JV
OpenAI has finalized a $10 billion joint venture with private equity firms. It raised over $4 billion to help businesses adopt its AI software. The initiative targets enterprise AI integration.
LLM Inference Hardware Crisis Worse Than Thought
Google DeepMind engineers' paper exposes how GPUs/TPUs, optimized for training, fail at memory-intensive LLM inference decoding, driving up costs. Trends like MoE and long contexts amplify the gap as memory bandwidth lags compute. They propose HBF, PNM, and 3D stacking to address core pain points.
10 Productivity Tips for M365 Mobile Apps
Article outlines mobile-first features in Microsoft 365 apps like Word for better smartphone productivity. Highlights Mobile View for easy reading with auto-adjust zoom and Headings navigation for quick document jumps. Notes Copilot AI availability but focuses on standard features for all users.
OpenClaw v2026.5.4-beta.1 Beta Released
OpenClaw has released version v2026.5.4-beta.1 as a beta update. The release is available on GitHub Releases. This marks a new beta iteration in the 2026.5.4 series.
Palantir Proves Resilience in Software Selloff
Palantir enters Q1 earnings under pressure to differentiate from software selloff. Investors question its inclusion in months-long market downturn. The company aims to demonstrate strong performance.
Semicon Giants Enter Hybrid Bonding for HBM
Leading storage, foundry, and equipment companies are aggressively entering the hybrid bonding market. This advanced packaging technology is poised for HBM memory, signaling imminent breakthroughs.