Daily Briefing

March 29, 2026

2026-03-28

35 articles

Less Gaussians, Texture More: 4K Feed-Forward Textured Splatting

2026-03-28

Summary

The LGTM framework proposed by Apple ML utilizes per-primitive textures to overcome the resolution scaling limits of existing feed-forward 3D Gaussian Splatting, enabling 4K novel view synthesis without per-scene optimization.

Key Points

Existing feed-forward 3DGS methods made 4K synthesis virtually impossible as the number of primitives increased quadratically with resolution.
LGTM combines compact Gaussian primitives with per-primitive textures to decouple geometric complexity from rendering resolution.
Achieved 4K high-quality novel view synthesis using a feed-forward method for the first time without per-scene optimization.
Maintains high-resolution rendering quality while using significantly fewer Gaussian primitives compared to existing methods.

Notable Quotes & Details

Notable Data / Quotes

First implementation of 4K resolution feed-forward novel view synthesis.
Solved the quadratic increase in the number of primitives.

Intended Audience

Computer vision and neural rendering researchers

Notes: The content is partially incomplete as it mixes content from unrelated previous research (HUGS 2023-12-07, Texturify 2022-10-05).

Anthropic's Claude popularity with paying consumers is skyrocketing

2026-03-28

Summary

The number of paid Claude subscribers from Anthropic is rapidly increasing in 2026.

Key Points

An Anthropic spokesperson stated that paid Claude subscriptions more than doubled in 2026.
Estimates for total Claude consumer users range from 18 million to 30 million, with variations across external analyses.
Anthropic has not officially released exact user data.

Notable Quotes & Details

Notable Data / Quotes

Paid subscriptions more than doubled (as of 2026)
Estimated total users: 18 million to 30 million

Intended Audience

AI industry professionals, investors, and competitor strategy officers

Notes: The source content is very short, and the summary is incomplete. No analysis of specific figures or reasons for growth is provided.

TikTok's policy for AI ads isn't working

2026-03-28

Summary

TikTok's AI-generated ad labeling policy is not effectively working, with Samsung cited as a primary example of non-compliance.

Key Points

Samsung placed an AI-generated Galaxy S26 Ultra ad on TikTok without a label, despite the same video being marked as AI-generated on YouTube.
TikTok's ad policy mandates labels (stickers, captions, watermarks, etc.) for content that is 'significantly modified' by AI.
Neither Samsung nor TikTok followed the policy, even though both are members of the Content Authenticity Initiative supporting C2PA standards.
Samsung did not respond to requests for comment, while TikTok only pointed to documentation regarding AI labeling requirements for advertisers.

Notable Quotes & Details

Notable Data / Quotes

TikTok policy definition: 'AI-significantly modified content' includes cases where the actions or speech of main subjects are altered by AI or the content is entirely AI-generated.

Intended Audience

General consumers, advertising industry professionals, and platform policy officers

Why OpenAI killed Sora

2026-03-28

Summary

An article analyzing the background behind OpenAI's termination of the video generation app Sora, citing excessive computing costs, intensifying competition, and investor skepticism as primary causes.

Key Points

Along with the termination of Sora, OpenAI also halted a $1 billion deal with Disney and withdrew plans for video generation features within ChatGPT.
Sora failed to secure an advantage in specific use cases compared to competing models from Kling, Google, etc.
There was a significant gap between Sora's launch marketing videos and the actual product, and download figures after launch were reportedly below expectations.
OpenAI is shifting focus toward business productivity and discontinuing 'side quests' to improve profitability.
Announcement of an additional $10 billion investment on the same day (total cumulative over $120 billion).

Notable Quotes & Details

Notable Data / Quotes

Halted $1 billion Disney deal
Additional $10 billion investment, total cumulative over $120 billion
Fidji Simo (OpenAI AGI Deployment CEO): "We cannot miss this moment because we are distracted by side quests."

Intended Audience

AI industry professionals, investors, and video generation AI developers

NVIDIA AI Unveils ProRL Agent: A Decoupled Rollout-as-a-Service Infrastructure for Reinforcement Learning of Multi-Turn LLM Agents at Scale

2026-03-28

Summary

NVIDIA unveiled ProRL AGENT, a distributed infrastructure for reinforcement learning of multi-turn LLM agents, significantly improving scalability by decoupling the rollout and training loops.

Key Points

The 'Rollout-as-a-Service' architecture separates rollout orchestration from the training loop, resolving conflicts between I/O-intensive environmental interaction and GPU-intensive training.
An asynchronous pipeline (INIT→RUN→EVAL) assigns each stage to independent worker pools to maximize throughput.
Adopted Singularity-based sandboxing instead of Docker, allowing execution without root privileges on Slurm-managed HPC clusters.
Replaced tmux-based terminal multiplexing with ptyprocess direct pseudo-terminals, reducing bash command latency from 0.78s to 0.42s.
Applied min-heap based load balancing to the vLLM inference backend pool to maximize prefix cache reuse.

Notable Quotes & Details

Notable Data / Quotes

Bash command latency: 0.78s → 0.42s (approx. 46% reduction)
Paper: https://arxiv.org/pdf/2603.18815
Comparison frameworks: SkyRL, VeRL-Tool, Agent Lightning, rLLM, GEM

Intended Audience

AI/ML researchers, reinforcement learning infrastructure engineers, and LLM agent developers

Show GN: 100% Autonomous Stock Trading System Linking 31 LLMs in a Cross-Verification System

2026-03-28

Summary

An article sharing a student developer's experience building and operating a fully autonomous stock trading system (K-Agent Alpha) composed of 31 LLM agents, linked to a real account with 10 million KRW.

Key Points

Operates a 5-stage Multi-Agent relay pipeline mimicking top-down investment: Macroeconomics, Industry, Corporate Analysis, Risk, and Final Decision.
Uses Gemini Flash for 30 out of 31 agents, and gemini-3.1-pro-preview for one CIO agent.
Solved the LLM 'disposition effect' (inability to stop losses) using Red-Teaming and past win-rate self-evaluation logic.
Stabilized token explosion and API rate limit issues with a batch processing architecture, keeping execution time under 1 hour.
Real account trading results and AI reports are streamed daily at 3:05 PM KST to a Telegram channel.

Notable Quotes & Details

Notable Data / Quotes

Linked to a real account with 10 million KRW
31 agents
Daily 3:05 PM Telegram channel disclosure

Intended Audience

Developers and readers interested in AI agent systems

Z.AI Coding Plan, GLM-5.1 Model Support — How to Switch from Claude Code/OpenClaw

2026-03-28

Summary

Z.AI's GLM Coding Plan has started supporting the latest GLM-5.1 model, and users can switch from existing coding agents like Claude Code by simply modifying configuration files.

Key Points

GLM-5.1 is now supported across all plans: Max, Pro, and Lite.
In Claude Code, you can switch by mapping Opus/Sonnet/Haiku environment variables to GLM models.
Existing agent workflows can be used as they are by simply modifying the configuration file.

Notable Quotes & Details

Intended Audience

Developers and AI coding tool users

Notes: Incomplete content — The source text was cut off in the middle.

Codex Plugin Released — Immediate Integration with Major Tools like Slack, Figma, Notion, Gmail

2026-03-28

Summary

OpenAI's coding agent Codex has released a plugin feature, evolving into a full-stack agent platform that can integrate with major collaboration tools such as Slack, Figma, Notion, and Gmail.

Key Points

Bundles Skills, Apps, and MCP servers into a single installation package for reusing workflows across teams and projects.
Provides an official plugin directory including GitHub, Slack, Notion, Linear, Gmail, Google Drive, Figma, and Vercel.
Supports local plugin scaffolding and automatic marketplace entry creation via the @plugin-creator skill.
Plugins are available across Codex App, CLI, and IDE extensions.
Expansion of support to tasks such as planning, research, and coordination prior to the coding stage.

Notable Quotes & Details

Intended Audience

Developers and software teams

Analysis of .claude/ Folder Structure

2026-03-28

Summary

An article analyzing and explaining the overall structure of the .claude/ folder (CLAUDE.md, commands/, skills/, agents/, settings.json, etc.), which is the core control directory of Claude Code.

Key Points

CLAUDE.md defines Claude's principles of behavior and project rules, with a recommendation to keep it under 200 lines.
Each Markdown file in the commands/ folder is automatically registered as a slash command (/), allowing shell command results to be inserted into prompts.
The skills/ folder defines workflows that are automatically triggered by analyzing conversation content using SKILL.md and YAML frontmatter.
The agents/ folder defines sub-agents with independent system prompts, models, and tool access permissions, realizing security and role separation.
settings.json controls command execution permissions and file access scope, while settings.local.json allows for personal overrides.

Notable Quotes & Details

Intended Audience

Claude Code users and developers

Collection of Codex Use Cases

2026-03-28

Summary

OpenAI released official documentation organizing 12 use cases where Codex can be applied in practice across 6 categories: Engineering, Front-end, Data, and more.

Key Points

Allows setting up automatic code reviews for GitHub PRs, with support for manual requests via @codex review comments.
Generates responsive UI code by reusing existing design system components and tokens from screenshots and design briefs.
Supports requesting full codebase explanations and key file recommendations when onboarding to unfamiliar repositories.
Enables score-based automatic iterative improvement loops when evaluation scripts are provided.
Provides 12 use cases across 6 categories: Engineering, Front-end, Data, Integrations, Mobile, and Evaluation.

Notable Quotes & Details

Notable Data / Quotes

12 use cases
6 categories

Intended Audience

Developers and software teams

[Project] PentaNet: Pushing beyond BitNet with Native Pentanary {-2, -1, 0, 1, 2} Quantization (124M, zero-multiplier inference)

2026-03-28

Summary

Announcement of the PentaNet project, which introduces pentanary quantization ({-2, -1, 0, +1, +2}) beyond BitNet's ternary quantization, achieving approx. 6.4% perplexity improvement while maintaining zero-multiplier inference benefits.

Key Points

±2 can be implemented with bit shifts (x<<1), allowing processing without hardware multipliers and maintaining BitNet's zero-multiplier advantage.
Pentanary quantization provides 47% more information than ternary (log₂(5)≈2.32 bits vs log₂(3)≈1.58 bits).
Achieved approx. 6.4% perplexity improvement at the same computing budget for 124M parameters based on WikiText-103.
Confirmed that ±2 buckets remain stable during training and do not collapse into ternary.
Released PyTorch PentaLinear layer implementation, NeurIPS-style technical paper, and weights as open-source on HuggingFace.

Notable Quotes & Details

Notable Data / Quotes

Approx. 6.4% perplexity improvement
124M parameters
log₂(5)≈2.32 bits vs log₂(3)≈1.58 bits (47% increase in information)
WikiText-103 benchmark

Intended Audience

AI/ML researchers and model quantization developers

[D] Thinking about augmentation as invariance assumptions

2026-03-28

Summary

An article discussing how to frame data augmentation as an invariance assumption and systematically reason about which transformations are valid and when they damage the training signal.

Key Points

Every augmentation is an invariance assumption, and transformations valid for one task can be destructive for another.
Depending on the intensity of the transformation, it can dilute the signal required by the model even if the label remains the same.
Discussion centers on computer vision examples, but the fundamental problem applies to broader domains.
Emphasis on the importance of verifying whether augmentation actually preserves labels.

Notable Quotes & Details

Intended Audience

ML researchers and data scientists

[R] Lag state in citation graphs: a systematic indexing blind spot with implications for lit review automation

2026-03-28

Summary

An article analyzing the structural blind spot of 'lag state' nodes in citation graphs—papers actively cited by recent works but not yet reflected in major indices—and discussing the impact on literature review automation pipelines.

Key Points

Lag state nodes are concentrated in papers that are rapidly being cited, causing automation pipelines to miss frontline materials.
A systematic gap exists in major index-based systems like Semantic Scholar.
Standard centrality metrics systematically underestimate cold nodes playing gateway, foundational, or protocol roles.
Direct bias impact on building citation graph embeddings and graph-based retrieval/search systems.
Released as a live research journal with 16+ entries in EMERGENCE_LOG.md.

Notable Quotes & Details

Notable Data / Quotes

16+ entries in EMERGENCE_LOG.md

Intended Audience

AI/ML researchers and literature review automation developers

Is anyone else watching what Qubic is doing with distributed compute and AI training? Seems underreported in AI circles

2026-03-28

Summary

An article introducing the 'Useful Proof of Work' method where the Qubic project conducts AI training (Aigarth AI) with distributed computing, asking if it is being seriously discussed in the AI research community.

Key Points

Qubic uses a 'Useful Proof of Work' method performing neural network training tasks instead of random hashes.
CertiK live mainnet verification recorded 15.52 million TPS (surpassing Visa's theoretical maximum).
Achieved high throughput with a bare-metal architecture without a VM layer.
DOGE mining integration scheduled for around April 1st (parallel operation of Scrypt ASICs and CPU/GPUs).
Differentiates itself from Bittensor in that mining itself is actual AI training.

Notable Quotes & Details

Notable Data / Quotes

15.52 million TPS (CertiK verified)
DOGE mining integration expected around April 1st

Intended Audience

Those interested in AI infrastructure and blockchain-AI convergence researchers

Notes: May be promotional; community posting containing several spelling errors.

Looking for a solid ChatGPT alternative for daily work

2026-03-28

Summary

An article sharing the experience of switching to a single unified platform to reduce subscription costs for multiple AI services (Claude, Gemini, GPT-4, etc.), using over 200 models in one place.

Key Points

Reduced spending from over $100/month for individual subscriptions by about half using a unified platform.
Access to over 200 models on a single platform.
Improved efficiency in coding and document review without usage limits or cooldowns.
Presents a case of processing a 100-page research paper with a long-context model.

Notable Quotes & Details

Notable Data / Quotes

Over $100/month → reduced by approx. half
200+ models

Intended Audience

Daily users of AI tools and those seeking cost reduction

Notes: Community posting that may promote a specific service.

Nobody's talking about what Pixar's Hoppers is actually saying about AI

2026-03-28

Summary

An article sharing an interpretation that the Pixar film 'Hoppers' allegorically depicts AI risks, alignment problems, and governance.

Key Points

The protagonist Dr. Sam's invention is compared to LLMs breaking human-machine communication barriers.
AI alignment problems, where unintended consequences occur due to the technology's own logic and momentum, are portrayed on screen.
Explicitly includes a governance message that a single individual or group should not control powerful technology.
The true warning is aimed at users who believe developers are the 'only solution,' rather than at the developers themselves.

Notable Quotes & Details

Intended Audience

General readers interested in AI philosophy and ethics

I have created a biologically based AI model

2026-03-28

Summary

An article introducing NIMCP, a biologically inspired artificial brain that simultaneously trains 6 types of neural networks and includes structural safety modules.

Key Points

SNNs achieved a 26Hz firing rate and 67% sparsity (within mammalian cortical range) without normalization (naturally occurring under cross-network training pressure).
Ethics modules are implemented as function calls in the inference code path rather than learned weights, making them impossible to fine-tune or jailbreak.
Uses curiosity-based learning (prediction error → dopamine → STDP gating) to learn without a separate reward function.
Currently in Stage 2 of a 4-stage developmental curriculum (Sensory → Naming → Feedback → Reasoning).
2,600 source files, 240 Python API methods, 8 language bindings, running on a single RTX 4000 (20GB VRAM) GPU.

Notable Quotes & Details

Notable Data / Quotes

26Hz firing rate, 67% sparsity
2,600 source files
RTX 4000 20GB VRAM

Intended Audience

AI/ML researchers and those interested in neuroscience and AI safety

I built a single platform integrating GPT-5.2, Grok 4, Claude 3.5, Gemini 3.1 Pro, Luma, Kling, ElevenLabs, OpenAI WebRTC and 50+ tools with shared persistent memory - is this the future of AI or have I over-engineered a mess?

2026-03-28

Summary

A post by a solo founder who self-taught for 3 months to build a multi-AI platform with 18 API integrations and persistent memory, asking the community if it's over-engineered or the future direction.

Key Points

Simultaneous integration of 18 APIs including OpenAI, Anthropic, Google, xAI, DeepSeek, Luma AI, Kling, and ElevenLabs.
Maintains full memory continuity using OpenAI Assistants API vector stores even when switching models during conversation.
Implemented a credit economy system and a dual payment architecture with Stripe web and Android/Apple IAP.
Over 50 features including real-time 2-way voice conversation (WebRTC), video/music generation, image editing, and code generation.
Vercel deployment, Firebase Firestore database, Sentry error tracking, and IPify IP rate limiting.

Notable Quotes & Details

Notable Data / Quotes

18 API integrations
Over 700 commits
Over 1,000 hours invested
3 months self-taught

Intended Audience

Developers and founders interested in building AI platforms

Notes: May be promotional; includes currently non-existent model names like GPT-5.2, Grok 4, and Gemini 3.1 Pro.

The AI releases hype cycle in a nutshell

2026-03-28

Summary

A critical observation that new AI feature announcements always follow the same pattern (enthusiasm in week 1 → disappointment in week 2) and that companies reset the cycle with new announcements.

Key Points

Week 1: Excessive excitement due to amazing demos of new models like VEO 3 or GPT-5.4.
Week 2: Disappointing reality experiences such as ignored prompts, flooding of em dashes, and quality degradation.
Companies shift focus to new feature announcements without disclosing quality degradation.
The hype cycle repeats, and users continue to be exposed to the same pattern.

Notable Quotes & Details

Intended Audience

General users of AI services and those interested in the AI industry

TurboQuant on MLX: 4.6x KV cache compression with custom Metal kernels (Qwen 32B at 98% FP16 speed)

2026-03-28

Summary

Sharing a project that implemented Google's TurboQuant KV cache compression for MLX, achieving 4.6x compression and 98% of FP16 speed with custom Metal kernels.

Key Points

Achieved 4.6x compression, 0.98x FP16 speed, and same quality on Qwen2.5-32B, M4 Pro 48GB.
KV cache reduced from 4.2GB to 897MB in 16K context.
Improved speed from 0.28x to 0.98x FP16 with fused Metal quantization/dequantization kernels and incremental decode buffers.
PR submitted to mlx-lm and released as open-source.

Notable Quotes & Details

Notable Data / Quotes

4.6x compression
16K context: 4.2GB → 897MB
M4 Pro 48GB

Intended Audience

Local LLM execution developers and Apple Silicon users

llama.cpp: Prefetching weights when offloading to CPU

2026-03-28

Summary

Sharing an experimental PR for llama.cpp that prefetches weights when offloading to CPU, improving prompt processing (PP) performance for dense and small MoE models.

Key Points

Weight prefetching during CPU offloading improves prompt processing performance.
Effective for dense models and small MoE models.
Suitable for 'ram-rich, gpu-poor' environments with abundant RAM but limited GPU.
Experimental PR (#21067) released on GitHub.

Notable Quotes & Details

Intended Audience

llama.cpp users and local LLM execution developers

M5 Max vs M3 Max Inference Benchmarks (Qwen3.5, oMLX, 128GB, 40 GPU cores)

2026-03-28

Summary

Sharing benchmark results comparing inference performance of 3 types of Qwen 3.5 models on M5 Max and M3 Max MacBook Pros (each with 40 GPU cores and 128GB unified memory).

Key Points

35B-A3B MoE: M5 Max 134.5 vs M3 Max 80.3 tok/s (1.7x difference).
122B-A10B MoE: M5 Max 65.3 vs M3 Max 46.1 tok/s (1.4x difference).
Based on 27B dense in 65K context: M5 Max 19.6 vs M3 Max 6.8 tok/s (2.9x), with prefill differing up to 4x.
MoE model efficiency: 122B model (10B active) is faster than 27B dense — speed is determined by active parameters, not total parameters.
M5 Max batch processing advantage: 2.54x throughput improvement with 4x batch on 35B-A3B.

Notable Quotes & Details

Notable Data / Quotes

M5 Max 614 GB/s vs M3 Max 400 GB/s memory bandwidth
Up to 2.9x inference speed advantage in 65K context

Intended Audience

Apple Silicon users and local LLM execution developers

Built a simple PyTorch flash-attention alternative for AMD GPUs that don't have it

2026-03-28

Summary

An article sharing the experience of creating a PyTorch-based tiling attention implementation to enable video generation on AMD GPUs (MI50/gfx906) that do not support Flash Attention.

Key Points

MI50 (gfx906) is excluded from official support in all major optimized attention implementations like CK, AOTriton, Flash Attention ROCm, and Triton.
Query dimension tiling reduced memory usage from O(N²) to O(N), decreasing from 26GB to ~1GB based on 17K tokens.
Implemented a 3-stage fallback structure: Standard chunk → Online Softmax → In-place Softmax.
Automatic conversion from BF16 to FP16 (since gfx906 lacks BF16 hardware support) and prevention of FP16 denormal NaN with FTZ threshold.
Completed a stable pure PyTorch implementation after 28 iterations.

Notable Quotes & Details

Notable Data / Quotes

28 iterations
MI50 gfx906
17K tokens (2.5s 480p video): Full attention score matrix 26GB
75K tokens (5s 720p): Full attention score matrix over 500GB

Intended Audience

AMD GPU users and local LLM/video generation developers

Google's Internal 'Agent Smith' Exploding in Popularity... "Even Restricting Access"

2026-03-28

Summary

Google is operating an internal coding agent 'Agent Smith,' which is so popular among employees that access must be restricted due to surging demand.

Key Points

A coding agent built on the agent-centric development platform 'Antigravity'.
Operates asynchronously in the background, allowing monitoring and direction via smartphone.
CEO Sundar Pichai stated that AI adoption would be reflected in performance evaluations.
Meta (Myclo) and Amazon are also developing similar internal agents.
Evolved from Google's 2024 internal coding model 'Goose,' it is expected to become a base for corporate agent development.

Notable Quotes & Details

Notable Data / Quotes

Sundar Pichai, Google CEO: Stated that AI adoption will be reflected in performance evaluations.
Sergey Brin: Announced that agents will play a major role at Google this year.

Intended Audience

Tech industry employees and corporate AI strategy officers

150 Trillion KRW in Memory Market Cap Vanishes Due to Google's 'TurboQuant' Shock... "A Second DeepSeek Incident?"

2026-03-28

Summary

The announcement of Google's 'TurboQuant' paper, which reduces LLM memory usage by 1/6th, caused a decrease of about $100 billion (151 trillion KRW) in the market capitalization of US memory chip-related stocks.

Key Points

Micron's market cap fell by over $70 billion (106 trillion KRW), a 15% drop; SanDisk lost $15 billion.
Professor Han In-soo from KAIST participated in the joint TurboQuant research.
Morgan Stanley remained neutral in the short term, mentioning potential long-term demand increase due to Jevons' Paradox.
Cloudflare's CEO evaluated TurboQuant as 'Google's DeepSeek moment'.
Sony raised the price of PlayStation 5 by up to 20% due to rising memory component prices.

Notable Quotes & Details

Notable Data / Quotes

Market cap of US memory chip stocks fell by about $100 billion (151 trillion KRW).
Micron dropped 15%, a $70 billion (106 trillion KRW) loss.
SanDisk lost $15 billion (22.6 trillion KRW) in market cap in one week.

Intended Audience

Investors, semiconductor industry employees, and tech industry stakeholders

Apple Recruits Google Veteran Executive as VP of AI Marketing to Improve 'Siri'

2026-03-28

Summary

Apple has recruited Lillian Rincon from Google as Vice President of AI Product Marketing to strengthen Siri and Apple Intelligence.

Key Points

Lillian Rincon led the global Google Assistant and Google Shopping organizations for 9 years at Google.
She will oversee product marketing and product management for Apple Intelligence and the Siri AI platform.
Plans to release a next-generation Siri with advanced AI within this year.
Exploring an open system where Google Gemini can be integrated into Siri and other third-party models can be selected.
Following the recruitment of Amar Subramanya from Google/MS as AI VP in December 2024, this marks continuous talent acquisition.

Notable Quotes & Details

Intended Audience

Tech industry employees and consumer technology enthusiasts

OpenAI Introduces Plugins Connecting Work Tools to 'Codex'... Responding to 'Claude Co-work'

2026-03-28

Summary

OpenAI added a plugin feature to its coding tool 'Codex' that integrates with external services like Slack, Google Drive, and Notion, expanding its work automation ecosystem.

Key Points

Automates repetitive tasks with workflow plugins bundling Skills, App integration, and MCP servers.
Released an official directory of over 20 plugins including Slack, Google Drive, Gmail, Figma, Notion, and GitHub.
Codex weekly active users surpassed 1.6 million, and a Windows version was released.
Plans to add a marketplace feature where users can distribute plugins themselves.
Standardization competition based on MCP alongside Anthropic's 'Claude Co-work' and Google's 'Gemini CLI' extensions.

Notable Quotes & Details

Notable Data / Quotes

Codex weekly active users surpassed 1.6 million.
Over 20 plugins released in the official directory.

Intended Audience

Developers and corporate IT officers

OpenAI's 'ChatGPT Ads' Surpass 150 Billion KRW in Revenue in 6 Weeks

2026-03-28

Summary

OpenAI's ChatGPT advertising pilot surpassed $100 million (approx. 150 billion KRW) in annualized revenue within 6 weeks of launch, showing faster growth than expected.

Key Points

Pilot advertising is running for free users and 'Go' low-price plan users in the US.
Over 600 advertisers are participating, with approx. 80% being small and medium-sized businesses.
Plans to introduce a self-serve platform for advertisers and expand to Canada, Australia, and New Zealand in April.
No ad intervention in the AI response generation process, and user conversation content is not shared with advertisers.
Ads are not shown to users under 18 or near sensitive topics like politics, health, and mental health.

Notable Quotes & Details

Notable Data / Quotes

Annualized revenue surpassed $100 million (approx. 150 billion KRW) in 6 weeks.
Over 600 advertisers, 80% of which are SMBs.
85% of US users are eligible for ad exposure, while the actual daily exposure rate is under 20%.

Intended Audience

Investors and AI business stakeholders

Datastreams: "Solving AI Limits with Data Architecture and Governance"

2026-03-28

Summary

Datastreams announced an enterprise AI strategy based on Data Fabric, arguing that AI limits should be solved through data architecture and governance systems rather than models.

Key Points

LLMs are difficult to apply directly to tasks where errors are not allowed, such as financial calculations, legal notices, or civil complaint handling.
Proposed a structure using Data Fabric to control LLMs so they only query 'correct' data.
Architecture combining metadata-based virtualization, data quality management, and Data Lineage.
Case study at Korea Expressway Corporation: Advanced AI as a practical work tool through data governance.
Proposal to build a 'National AI Compliance Service System' based on LLM/RAG for small and medium-sized enterprises.

Notable Quotes & Details

Notable Data / Quotes

"What data AI is designed to judge based on is more important than how smart AI is" — Lee Young-sang, CEO of Datastreams

Intended Audience

Corporate AI officers, data engineers, and public institution IT officers

Notes: Seminar report containing corporate promotional presentation content.

[AI Now] Repeated Claude Outages Amid Surging Demand... Anthropic's Policy and Infrastructure Under Test

2026-03-28

Summary

Anthropic is under test on both policy and infrastructure fronts as users surge due to conflicts with the Trump administration while service outages repeat.

Key Points

Failure of contract renewal with the US War Department: Refused to delete clauses related to fully autonomous weapons and domestic surveillance, resulting in a $200 million contract cancellation.
President Trump ordered federal agencies to immediately stop using Claude and designated Anthropic as a 'supply chain risk' company.
Claude App ranked #1 in the App Store after the conflict was publicized, with free users increasing by over 60% and paid subscribers doubling since January.
Repeated service outages, including a total suspension of the consumer app on the 2nd of this month and complete login blocking on the 22nd.
The California Federal Court suspended the effectiveness of the supply chain risk designation and usage ban with a preliminary injunction.

Notable Quotes & Details

Notable Data / Quotes

Free users increased by over 60% compared to January.
Paid subscribers more than doubled since October 2025.
Daily new signups quadrupled.
Canceled $200 million contract with the War Department.
Over 2,000 outage reports on Downdetector around 6:40 AM.

Intended Audience

AI industry employees, policy stakeholders, and general readers

[Kang Eun-sung Security Column ④] Living with Artificial Intelligence — Literacy, Questions, and Judgment

2026-03-28

Summary

A column by a security expert pointing out the risks of over-reliance in the AI era and the importance of literacy, proper questions, and judgment.

Key Points

Concerns about over-reliance as elementary school students regard AI as emotional friends.
Included 'Over-reliance' as the 9th vulnerability in OWASP Top 10 for LLM v1.0 (August 2023).
LLMs do not speak facts but predict the most probable words based on probability.
Rapid evolution into multimodal, agentic AI, and physical AI in about 3 years since ChatGPT's launch.
Emphasis on the need for AI literacy for the 2030 youth generation and the AI 'immigrant' generation.

Notable Quotes & Details

Notable Data / Quotes

OWASP Top 10 for LLM v1.0 (Announced August 2023)
OWASP Top 10 for LLM applications 2025 (Announced November 2024, re-categorizing over-reliance as a sub-element of 'misinformation')

Intended Audience

General readers, educators, and parents

Notes: Column-style content centered on opinions and perspectives; the main text appears to be cut off in the middle.

"OLED Monitor Shipments to Rise 51% This Year" — TrendForce

2026-03-28

Summary

Market research firm TrendForce predicted that OLED monitor shipments in 2026 will reach approx. 4.13 million units, a 51% increase from the previous year.

Key Points

2026 OLED monitor shipments expected to be approx. 4.13 million units (51% increase from 2.735 million in 2025).
ASUS ranks 1st with 21.6% share, Samsung Electronics 2nd with 19.3%, MSI 3rd with 13.1%, and LG Electronics 4th with 12.6%.
Demand driven by the cost-effectiveness of 27-inch 240Hz QHD OLED monitors, with new 280Hz products also having a positive impact.
Samsung Display (QD-OLED) and LG Display (W-OLED) lead panel supply.
AVC Revo predicts higher OLED monitor shipments this year at 5.4 million units.

Notable Quotes & Details

Notable Data / Quotes

2026 OLED monitor shipments expected to be approx. 4.13 million units (up 51% YoY).
ASUS 1st with 21.6% share (surpassed Samsung Electronics in Q3 2025 to take 1st place).
AVC Revo forecast: Samsung Display 4 million units, LG Display 1.4 million units, totaling 5.4 million units.

Intended Audience

Consumers, IT hardware enthusiasts, and investors

TA446 Deploys DarkSword iOS Exploit Kit in Targeted Spear-Phishing Campaign

2026-03-28

Summary

The Russian FSB-linked threat group TA446 deployed a spear-phishing campaign targeting a wide range of subjects including government, think tanks, and finance, utilizing the leaked DarkSword iOS exploit kit.

Key Points

TA446 (Callisto/COLDRIVER/Star Blizzard) is a Russian FSB-linked state-sponsored hacking group previously focused on credential theft and WhatsApp account attacks.
In this campaign, they distributed GHOSTBLADE data theft malware via the DarkSword exploit kit through 'discussion invitation' emails impersonating the Atlantic Council.
The DarkSword kit consists of an initial redirector, exploit loader, remote code execution (RCE), and PAC (Pointer Authentication Code) bypass components.
MAYBEROBOT backdoor is also distributed via password-protected ZIP files, and attack volume has 'significantly increased' in the last 2 weeks.
Targets are much broader than before, including government, think tanks, higher education, finance, and legal institutions, as well as Russian opposition politician Leonid Volkov.

Notable Quotes & Details

Notable Data / Quotes

"We have not previously observed TA446 target users' iCloud accounts or Apple devices, but the adoption of the leaked DarkSword iOS exploit kit has now enabled the actor to target iOS devices" — Proofpoint
Email sent date: 2026-03-26
Second-level domain referenced by DarkSword loader: escofiringbijou[.]com

Intended Audience

Security researchers, corporate security officers, and threat intelligence analysts

Explanation for why we don't see two-foot-long dragonflies anymore fails

2026-03-28

Summary

The 'oxygen constraint hypothesis,' considered the standard theory for the extinction of ancient giant insects for 30 years, has been refuted by recent research.

Key Points

In the late Paleozoic era approx. 300 million years ago, giant insects like Meganeuropsis permiana with a 70cm wingspan and 100g weight existed.
The existing 'oxygen constraint hypothesis' explained that giant insects could not survive when atmospheric oxygen levels decreased due to the inefficient tracheal respiration method of insects.
Edward Snelling, a professor of veterinary science at the University of Pretoria, argued this hypothesis is wrong.
Insects lack lungs and a circulatory system, unlike mammals, birds, and reptiles, and breathe through a tracheal system.
The 'simple and elegant explanation' used for 30 years has been proven wrong, requiring a new explanation.

Notable Quotes & Details

Notable Data / Quotes

Wingspan of Meganeuropsis permiana: over 70cm, Weight: 100g
"It's a simple, elegant explanation, but it's wrong." — Edward Snelling, Professor of Veterinary Science at the University of Pretoria

Intended Audience

General readers in science and biology

Notes: The source content includes only the article introduction, and details of the refutation are cut off.

Best Amazon Spring Sale deals under $25

2026-03-28

Summary

Introducing a list of cost-effective tech products available for under $25 at the Amazon Big Spring Sale.

Key Points

Amazon Fire TV Stick: Includes Alexa Voice Remote, can control smart home devices.
MagSafe Power Bank 5,000mAh/18Wh: Ultra-slim design with dimensions 3.9×2.6×0.3 inches and weight 3.8oz.
1080p HD Indoor Smart Security Camera: Motion detection and 2-way audio, 50% discount.
30oz Water Bottle: Introduced as the most frequently used item by the shopping editor.
Includes various household appliances like a small portable vacuum cleaner.

Notable Quotes & Details

Notable Data / Quotes

Power bank capacity: 5,000mAh/18Wh
Security camera discount rate: 50% off
Power bank weight: 3.8oz (approx. 108g)

Intended Audience

General consumers looking for cost-effective tech products

Notes: Commercial recommendation article based on affiliate marketing; the latter half of the article is cut off.

PreviousDaily Briefing

NextDaily Briefing