# forck.live — 3–9 March 2025

> AI lab announcements, news reports and podcasts.

forck.live follows what companies announce about their own products. Press may
corroborate an official story; it never leads and never appears as an item of its
own. A missing date stays missing. A title is the publisher's words — where a
company does not publish in English, an English rendering is listed separately
as `Title in English` and named as this site's translation, never as theirs.

This file is one issue of the weekly archive: every confirmed
first-party announcement of ISO week 2025-W10, 3–9 March 2025, newest
first. The window never moves, so this URL always names the same seven days —
but a post published inside them and detected later is added when it arrives,
so a closed week can still gain entries.

Every entry is one first-party announcement, and carries the same eight fields
in this order before any others:

    Company · Announced · Category · Coverage · Announcement · Group · Source · Record

`Source` is the company's own page the announcement was traced to and `Record`
is this site's page for it. Both are safe to cite.

`Coverage` is a number of outlets, or `not counted` — which means this record
reached no count of its own, never that nobody wrote about it.

`Category` is this tracker's own analysis, and is authoritative wherever it is
stated. `not stated` means either that nothing has analysed the post yet or
that the analysis placed it outside this taxonomy; neither is a verdict about
the announcement.

`Announcement` is whether the post is news of something the company did, rather
than a how-to, a customer story, a changelog listing or a staffing note.

`Group` is which of this record's groups the entry falls in, carried on the
entry itself so it holds where the groups are headings and where they are not.
One of: `models`, `covered`, `announcements`, `routine`.

Anything after `Record` is held detail that only some announcements carry.

- This week's page: https://forck.live/week/2025-W10
- The current issue, and the catalogue of the other weeks: https://forck.live/briefing.md
- Catalogue: https://forck.live/llms.txt

## Announcements

### Replit — Everything you need to know about MCP

- Company: Replit (replit.com)
- Announced: 2025-03-09T01:05:24+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://replit.com/blog/everything-you-need-to-know-about-mcp
- Record: https://forck.live/items/17953-everything-you-need-to-know-about-mcp
- Subject: Replit Agent

Replit's post explains the Model Context Protocol (MCP), a standard for connecting AI models to external data sources and tools. It describes MCP's client-server architecture, its benefits such as ready-to-use integrations and provider switching, and provides a quickstart template for developers.

### Vercel — Personalization strategies that power ecommerce growth

- Company: Vercel (vercel.com)
- Announced: 2025-03-07T13:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://vercel.com/blog/personalization-strategies-that-power-ecommerce-growth
- Record: https://forck.live/items/20896-personalization-strategies-that-power-ecommerce-growth
- Subject: v0 / AI SDK / AI Gateway

Vercel's guide covers two core types of personalization—architectural and content—and explains how Next.js and Vercel tools such as Edge Middleware, Partial Prerendering, and the Flags SDK enable performance-focused personalization for ecommerce. It notes that effective personalization can return $20 for every $1 spent and drive 40% more revenue, while cautioning against common mistakes like over-use of client-side personalization and using configuration tools to manage content.

### OpenAI — Nubank elevates customer experiences with OpenAI

- Company: OpenAI (openai.com)
- Announced: 2025-03-07T08:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://openai.com/index/nubank
- Record: https://forck.live/items/609-nubank-elevates-customer-experiences-with-openai
- Subject: GPT / ChatGPT / API

Nubank is improving customer experiences using OpenAI.

### Supabase — Dedicated Poolers

- Company: Supabase (supabase.com)
- Announced: 2025-03-07
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://supabase.com/blog/dedicated-poolers
- Record: https://forck.live/items/21373-dedicated-poolers
- Subject: Supabase

Supabase announced Dedicated Poolers, a PgBouncer instance co-located with the user's Postgres database for maximum performance and reliability. It is available for select customers now and will be generally available by March 20, 2025. The Dedicated Pooler is isolated to the user's project and offers fine-grained configuration control, available on the Pro Plan and above.

### Upstage — Why table structure extraction fails: A deep dive into real-world challenges

- Company: Upstage (upstage.ai)
- Announced: 2025-03-07
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.upstage.ai/blog/en/why-table-structure-extraction-fails-a-deep-dive-into-real-world-challenges
- Record: https://forck.live/items/18474-why-table-structure-extraction-fails-a-deep-dive-into-real-world-challenges
- Subject: Solar Pro / Solar Mini

Upstage published a technical comparison evaluating table structure extraction across document-parsing products, focusing on merged cells, multi-section layouts, and hierarchical tables. The post states that its Document Parse tool retains merged cells, preserves table formatting in HTML, and accurately extracts tables from scanned PDFs and multi-section layouts while maintaining structural integrity. It acknowledges some limitations in maintaining hierarchical header relationships.

### LG AI Research — The Story of Ahra Jo, Leading the DDU Research to Build an AI That Understands All Documents in the World

- Company: LG AI Research (lgresearch.ai)
- Announced: 2025-03-07
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.lgresearch.ai/blog/view?seq=538
- Record: https://forck.live/items/4801-the-story-of-ahra-jo-leading-the-ddu-research-to-build-an-ai-that-understands
- Subject: EXAONE

LG AI Research's Deep Document Understanding (DDU) technology enables AI to comprehend various types of information in documents, including pictures, graphs, and tables. The technology has been partially applied to ChatEXAONE, an AI assistant for enterprise professionals, and is used in the Financial AI project and EXAONE Discovery platform.

### Hugging Face — LLM Inference on Edge: A Fun and Easy Guide to run LLMs via React Native on your Phone!

- Company: Hugging Face (huggingface.co)
- Announced: 2025-03-07
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://huggingface.co/blog/llm-inference-on-edge
- Record: https://forck.live/items/1735-llm-inference-on-edge-a-fun-and-easy-guide-to-run-llms-via-react-native-on
- Subject: Platform

This is a tutorial blog post that guides readers through creating a React Native mobile app for running LLMs locally on a phone, using llama.rn and llama.cpp to load GGUF files from Hugging Face. It discusses model selection, quantization formats, and environment setup.

### Harvey — Harvey's Values

- Company: Harvey (harvey.ai)
- Announced: 2025-03-06T17:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.harvey.ai/blog/harvey-values
- Record: https://forck.live/items/17714-harvey-s-values
- Subject: Harvey

Harvey announced its three company values: Decisiveness, Simplicity, and Job's Not Finished (JNF), intended to guide operations, collaboration, and culture as the company grows.

### OpenAI — Accelerating engineering cycles 20% with OpenAI

- Company: OpenAI (openai.com)
- Announced: 2025-03-06T09:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://openai.com/index/factory
- Record: https://forck.live/items/610-accelerating-engineering-cycles-20-with-openai
- Subject: GPT / ChatGPT / API

The source text states that OpenAI is accelerating engineering cycles by 20%.

### Replit — Using Tvix Store to Reduce Nix Storage Costs by 90%

- Company: Replit (replit.com)
- Announced: 2025-03-06T04:46:22+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://replit.com/blog/tvix-store
- Record: https://forck.live/items/17981-using-tvix-store-to-reduce-nix-storage-costs-by-90
- Subject: Replit Agent

Replit implemented Tvix Store, a content-addressed Nix store implementation, to compress its 6TB cache of Nix store paths into 1.2TB, reducing storage costs for the persistent disk cache by 90%. The implementation uses a FUSE filesystem to expose the compressed store while maintaining backward compatibility for older packages.

### Amp — Raising an Agent, Episode 1

- Company: Amp (ampcode.com)
- Announced: 2025-03-06
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://ampcode.com/podcast
- Record: https://forck.live/items/7855-raising-an-agent-episode-1
- Subject: Amp

### Mistral AI — Mistral OCR

- Company: Mistral AI (mistral.ai)
- Announced: 2025-03-06
- Category: product-launch
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://mistral.ai/news/mistral-ocr
- Record: https://forck.live/items/3329-mistral-ocr
- Subject: Mistral / Le Chat

By Mistral AI Team

### Alibaba — QwQ-32B: Embracing the Power of Reinforcement Learning

- Company: Alibaba (alibaba.com)
- Announced: 2025-03-05T16:00:04+00:00
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://qwenlm.github.io/blog/qwq-32b/
- Record: https://forck.live/items/2294-qwq-32b-embracing-the-power-of-reinforcement-learning
- Subject: Qwen

Alibaba's Qwen team discusses the potential of reinforcement learning to improve model performance, citing examples like DeepSeek R1.

### Cartesia — Llamba: scaling distilled recurrent models for efficient language processing

- Company: Cartesia (cartesia.ai)
- Announced: 2025-03-05
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.cartesia.ai/blog/llamba-distillation
- Record: https://forck.live/items/8023-llamba-scaling-distilled-recurrent-models-for-efficient-language-processing
- Subject: Sonic / Ink

Cartesia released a technical report introducing MOHAWK, a multi-stage architecture distillation method that converts Transformer models into efficient Mamba-2 variants. The company released model weights for the Llamba family (Llamba-1B, Llamba-3B, Llamba-8B), distilled from corresponding Llama-3.X models, and reported throughput improvements up to 12× for Llamba-8B versus Llama-3.1-8B. The work targets on-device and high-throughput inference scenarios.

### Replicate — Wan2.1 parameter sweep

- Company: Replicate (replicate.com)
- Announced: 2025-03-05
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://replicate.com/blog/wan-21-parameter-sweep
- Record: https://forck.live/items/2502-wan2-1-parameter-sweep
- Subject: Platform

A parameter sweep experiment on Alibaba's WAN2.1 text-to-video model, testing guidance scale and shift parameters to understand their effects on video output. It provides recommendations for optimal values.

### Replicate — Wan2.1: generate videos with an API

- Company: Replicate (replicate.com)
- Announced: 2025-03-05
- Category: api-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://replicate.com/blog/wan-21-generate-videos-with-an-api
- Record: https://forck.live/items/2501-wan2-1-generate-videos-with-an-api
- Subject: Platform

Replicate announces that Wan2.1, an open-source video model, is now available via their cloud API, with multiple model flavors (480p, 720p, 1.3B) and code snippets for text-to-video and image-to-video generation.

### OpenAI — LaunchDarkly's approach to AI-powered product management

- Company: OpenAI (openai.com)
- Announced: 2025-03-04T10:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://openai.com/index/launchdarkly-claire-vo
- Record: https://forck.live/items/611-launchdarkly-s-approach-to-ai-powered-product-management
- Subject: GPT / ChatGPT / API

The post is a conversation with LaunchDarkly's CPO about product management and AI-native teams, not an announcement of a new AI product or model.

### OpenAI — Introducing NextGenAI

- Company: OpenAI (openai.com)
- Announced: 2025-03-04T06:00:00+00:00
- Category: funding-company-news
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://openai.com/index/introducing-nextgenai
- Record: https://forck.live/items/612-introducing-nextgenai
- Subject: GPT / ChatGPT / API

OpenAI announces a $50 million commitment in funding and tools to leading research institutions as part of the NextGenAI initiative.

### Factory AI — Building AI for Safe and Responsible Autonomy: Our Readiness Policy

- Company: Factory AI (factory.com)
- Announced: 2025-03-04
- Category: safety-policy-update
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://factory.com/news/safe-autonomy-readiness-policy
- Record: https://forck.live/items/17868-building-ai-for-safe-and-responsible-autonomy-our-readiness-policy
- Subject: Droid

Factory published its Safe Autonomy Readiness Policy (SARP), a governance framework for AI systems that can autonomously write, modify, or deploy production code. The policy defines frontier-level capabilities that trigger additional risk evaluation and safeguards, and outlines a four-step risk management process: identify, evaluate, mitigate, and monitor. Factory also established an internal Safety Review for frontier-level systems and commits to regular policy updates as AI capabilities advance.

### Mistral AI — Empowering product development with an agentic workflow

- Company: Mistral AI (mistral.ai)
- Announced: 2025-03-04
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://mistral.ai/news/agentic-workflows-from-meetings-to-dev-tickets
- Record: https://forck.live/items/3330-empowering-product-development-with-an-agentic-workflow
- Subject: Mistral / Le Chat

Mistral AI announces a solution for empowering product development with an agentic workflow.

### Hugging Face — A Deepdive into Aya Vision: Advancing the Frontier of Multilingual Multimodality

- Company: Hugging Face (huggingface.co)
- Announced: 2025-03-04
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://huggingface.co/blog/aya-vision
- Record: https://forck.live/items/1737-a-deepdive-into-aya-vision-advancing-the-frontier-of-multilingual-multimodality
- Subject: Platform

Cohere For AI announces Aya Vision, a new open-weight multilingual multimodal model family with 8B and 32B parameter variants, designed for vision-language understanding across 23 languages. The models outperform larger models in pairwise comparisons and are released as open weights for research.

### Hugging Face — Hugging Face and JFrog partner to make AI Security more transparent

- Company: Hugging Face (huggingface.co)
- Announced: 2025-03-04
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://huggingface.co/blog/jfrog
- Record: https://forck.live/items/1736-hugging-face-and-jfrog-partner-to-make-ai-security-more-transparent
- Subject: Platform

Hugging Face announces a partnership with JFrog to integrate JFrog's scanner into the Hugging Face Hub to improve AI security by reducing false positives and detecting malicious code in model weights.

### Lovable — Introducing Versioning 2.0 to Lovable

- Company: Lovable (lovable.dev)
- Announced: 2025-03-03T19:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://lovable.dev/blog/versioning-with-lovable-two-point-zero
- Record: https://forck.live/items/12149-introducing-versioning-2-0-to-lovable
- Subject: Lovable / AI app builder

Lovable released Versioning V2.0, a major update to its versioning system. The update introduces bookmarks for stable versions, a more intuitive version history that groups edits by date, and smarter restores that create new edit cards when reverting to old edits.

### Lovable — What Is Vibe Coding? A Beginner’s Guide to AI-First Development

- Company: Lovable (lovable.dev)
- Announced: 2025-03-03T14:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://lovable.dev/blog/what-is-vibe-coding
- Record: https://forck.live/items/12150-what-is-vibe-coding-a-beginner-s-guide-to-ai-first-development
- Subject: Lovable / AI app builder

Lovable's guide explains vibe coding, a method where users describe software requirements in natural language and AI generates the code. The post covers the origin of the term, how the process works, and a livestream demonstration where an event management app was built in over an hour without manually writing code.

### Vercel — Using the AI SDK to build Sitecore Stream's AI-powered brand aware assistant

- Company: Vercel (vercel.com)
- Announced: 2025-03-03T13:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://vercel.com/blog/using-the-ai-sdk-to-build-sitecore-streams-ai-powered-brand-aware-assistant
- Record: https://forck.live/items/20898-using-the-ai-sdk-to-build-sitecore-stream-s-ai-powered-brand-aware-assistant
- Subject: v0 / AI SDK / AI Gateway

Sitecore launched Sitecore Stream, an AI-powered brand assistant built using Vercel's AI SDK. The assistant allows marketers to interact with brand content conversationally, with real-time streaming and server-side rendering. The AI SDK enabled integration of Python backend with Next.js frontend, handling data streaming between Azure-hosted backends and the user interface.

### Vercel — How Fluid compute works on Vercel

- Company: Vercel (vercel.com)
- Announced: 2025-03-03T13:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://vercel.com/blog/how-fluid-compute-works-on-vercel
- Record: https://forck.live/items/20897-how-fluid-compute-works-on-vercel
- Subject: v0 / AI SDK / AI Gateway

Vercel introduced Fluid compute, a next-generation compute model that dynamically adjusts to traffic demands, reuses existing resources, and minimizes cold starts. The Vercel Functions router orchestrates execution to optimize resource usage, and projects with Fluid enabled have cut compute costs by up to 85%.

### Genspark — Announcing $100M Series A Funding and the First-Ever AI Search+Browser+Agents Super App!

- Company: Genspark (genspark.ai)
- Announced: 2025-03-03
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.genspark.ai/blog/genspark-series-a-funding-and-ios-app
- Record: https://forck.live/items/18293-announcing-100m-series-a-funding-and-the-first-ever-ai-search-browser-agents
- Subject: Genspark

### Cartesia — How to build a voice AI agent with Cartesia

- Company: Cartesia (cartesia.ai)
- Announced: 2025-03-03
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.cartesia.ai/blog/how-to-build-a-voice-ai-agent-with-cartesia
- Record: https://forck.live/items/8024-how-to-build-a-voice-ai-agent-with-cartesia
- Subject: Sonic / Ink

Cartesia published a tutorial on building voice AI agents using its API and the Sonic model. The guide covers setting up speech-to-text, response generation, and text-to-speech with Cartesia's TTS model sonic-3.5.