# forck.live — 23–29 March 2026

> 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 2026-W13, 23–29 March 2026, 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/2026-W13
- The current issue, and the catalogue of the other weeks: https://forck.live/briefing.md
- Catalogue: https://forck.live/llms.txt

## Announcements

### OpenAI — Helping disaster response teams turn AI into action across Asia

- Company: OpenAI (openai.com)
- Announced: 2026-03-29T22:15:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://openai.com/index/helping-disaster-response-teams-asia
- Record: https://forck.live/items/217-helping-disaster-response-teams-turn-ai-into-action-across-asia
- Subject: GPT / ChatGPT / API

A workshop on AI for disaster response in Asia, organized by OpenAI and the Gates Foundation.

### Alibaba — Qwen3.5-Omni: Scaling Up, Toward Native Omni-Modal AGI

- Company: Alibaba (alibaba.com)
- Announced: 2026-03-29T20:00:00+00:00
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://qwen.ai/blog?id=qwen3.5-omni
- Record: https://forck.live/items/4600-qwen3-5-omni-scaling-up-toward-native-omni-modal-agi
- Subject: Qwen

Alibaba's Qwen announces Qwen3.5-Omni, a new fully omnimodal LLM supporting text, images, audio, and audio-visual content. It comes in three sizes: Plus, Flash, and Light, with 256k long-context input. It is natively pretrained on massive data and offers enhanced multilingual capabilities compared to Qwen3-Omni. The model is available via Offline and Realtime APIs, and features include audio-visual captioning, voice cloning, and ARIA technique for speech stability.

### Google DeepMind — Reimagining the mouse pointer for the AI era

- Company: Google DeepMind (deepmind.google)
- Announced: 2026-03-29T10:50:49+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://deepmind.google/blog/ai-pointer/
- Record: https://forck.live/items/1243-reimagining-the-mouse-pointer-for-the-ai-era
- Subject: Gemini

Google DeepMind is introducing a context-aware AI pointer that aims to reduce the need for traditional prompting in Chrome and other contexts.

### OpenAI — STADLER reshapes knowledge work at a 230-year-old company

- Company: OpenAI (openai.com)
- Announced: 2026-03-27T22:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://openai.com/index/stadler
- Record: https://forck.live/items/218-stadler-reshapes-knowledge-work-at-a-230-year-old-company
- Subject: GPT / ChatGPT / API

STADLER uses ChatGPT to transform knowledge work, saving time and accelerating productivity across 650 employees.

### Harvey — Find the Case. Draft your Argument. All in Harvey.

- Company: Harvey (harvey.ai)
- Announced: 2026-03-27T19:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.harvey.ai/blog/new-us-case-law-in-harvey
- Record: https://forck.live/items/17588-find-the-case-draft-your-argument-all-in-harvey
- Subject: Harvey

Harvey added US case law coverage to its Harvey Knowledge platform, including opinions from the US Supreme Court, federal circuit and district courts, and state supreme, appellate, and trial courts. The new US case law knowledge source joins existing US legal sources covering legislation, regulation, official guidance, and primary law, alongside over 500 global sources. The offering will be available to customers in the coming weeks.

### Hugging Face — Liberate your OpenClaw

- Company: Hugging Face (huggingface.co)
- Announced: 2026-03-27
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://huggingface.co/blog/liberate-your-openclaw
- Record: https://forck.live/items/1530-liberate-your-openclaw
- Subject: Platform

Hugging Face publishes a guide for users of OpenClaw agents who have been cut off from Anthropic's Claude models, recommending open models such as GLM-5 and Qwen3.5-35B-A3B via Hugging Face Inference Providers or local setup with llama.cpp.

### Google DeepMind — Gemini 3.1 Flash Live: Making audio AI more natural and reliable

- Company: Google DeepMind (deepmind.google)
- Announced: 2026-03-26T15:23:35+00:00
- Category: model-update
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://deepmind.google/blog/gemini-3-1-flash-live-making-audio-ai-more-natural-and-reliable/
- Record: https://forck.live/items/1244-gemini-3-1-flash-live-making-audio-ai-more-natural-and-reliable
- Subject: Gemini

Google DeepMind's latest voice model has improved precision and lower latency to make voice interactions more fluid, natural and precise.

### Harvey — Turn Legal Analysis Into Deliverables Without Leaving Harvey

- Company: Harvey (harvey.ai)
- Announced: 2026-03-26T15:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.harvey.ai/blog/create-and-edit-files-in-harvey
- Record: https://forck.live/items/17589-turn-legal-analysis-into-deliverables-without-leaving-harvey
- Subject: Harvey

Harvey added the ability to generate PowerPoint presentations, Excel spreadsheets, and edit multiple Word documents within the platform, using its legal intelligence and firm templates. Users can now complete the entire workflow from analysis to export-ready deliverables without leaving Harvey.

### Supabase — Navigating Regional Network Blocks

- Company: Supabase (supabase.com)
- Announced: 2026-03-26
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://supabase.com/blog/navigating-regional-network-blocks
- Record: https://forck.live/items/21297-navigating-regional-network-blocks
- Subject: Supabase

Supabase shared lessons from three regional network blocks affecting its platform in the UAE, Yemen, and India between September 2025 and March 2026. The blocks were triggered by a single abusive project on the multi-tenant platform, and Supabase worked with authorities and customers to resolve them. The post explains how DNS-level and IP-level blocking works and notes that adding a domain to the Public Suffix List is a mitigation but not a complete prevention.

### Perceptron — Perceptron Vision MCP Server

- Company: Perceptron (perceptron.inc)
- Announced: 2026-03-26
- Category: developer-tool-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.perceptron.inc/blog/mcp
- Record: https://forck.live/items/18616-perceptron-vision-mcp-server
- Subject: Perceptron Isaac / Perceptron Mk1

Perceptron released an open-source MCP server that lets AI agents perform vision tasks—VQA, OCR, object detection, and captioning—using its Isaac model family, without routing every task through expensive frontier models.

### Krea — Annotations in Krea Edit

- Company: Krea (krea.ai)
- Announced: 2026-03-26
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.krea.ai/blog/annotations
- Record: https://forck.live/items/15084-annotations-in-krea-edit
- Subject: Krea 2 / AI creative platform

Krea Edit now supports annotations, allowing users to mark up multiple regions of an image, write separate prompts for each, and generate all changes in a single pass instead of sequentially. Each region can have its own prompt and optional reference image, and the sidebar lists all annotations for easy navigation.

### Suno — Suno v5.5: More Expressive. More You.

- Company: Suno (suno.com)
- Announced: 2026-03-26
- Category: model-update
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://suno.com/blog/v5-5
- Record: https://forck.live/items/10326-suno-v5-5-more-expressive-more-you
- Subject: Music / Studio

Suno released v5.5, its latest music-generation model, alongside three personalization features: Voices, which lets Pro and Premier subscribers capture and use their own singing voice; Custom Models, which allows Pro and Premier subscribers to fine-tune the model on their own tracks; and My Taste, a free feature that learns a user's genre and mood preferences over time.

### Amp — GPT‐5.4 in Deep

- Company: Amp (ampcode.com)
- Announced: 2026-03-26
- Category: model-update
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://ampcode.com/news/gpt-5.4
- Record: https://forck.live/items/7753-gpt-5-4-in-deep
- Subject: Amp

Amp announced that GPT-5.4 now powers its deep mode. The model was tuned to behave like GPT-5.3-Codex and is faster and better at coding. Users can access it via the command palette or editor extension, with options for deep^2 and deep^3 reasoning levels.

### Cohere — Introducing Cohere Transcribe: a new state-of-the-art in open-source speech recognition

- Company: Cohere (cohere.com)
- Announced: 2026-03-26
- Category: product-launch
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://cohere.com/blog/transcribe
- Record: https://forck.live/items/3269-introducing-cohere-transcribe-a-new-state-of-the-art-in-open-source-speech
- Subject: Command

Cohere introduces Cohere Transcribe, described as a new state-of-the-art in open-source speech recognition.

### Together AI — Plan, divide, and conquer: How weak models excel at long context tasks

- Company: Together AI (together.ai)
- Announced: 2026-03-26
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.together.ai/blog/plan-divide-conquer
- Record: https://forck.live/items/2366-plan-divide-and-conquer-how-weak-models-excel-at-long-context-tasks
- Subject: Inference platform

Together AI presents research showing that smaller models using a strategic Divide & Conquer framework can match or beat GPT-4o single-shot on long context tasks, with benefits of lower cost, faster processing, and easier tuning.

### Google DeepMind — Protecting people from harmful manipulation

- Company: Google DeepMind (deepmind.google)
- Announced: 2026-03-25T16:46:20+00:00
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://deepmind.google/blog/protecting-people-from-harmful-manipulation/
- Record: https://forck.live/items/1245-protecting-people-from-harmful-manipulation
- Subject: Gemini

Google DeepMind conducts research on harmful manipulation risks of AI in areas such as finance and health, leading to new safety measures.

### Google DeepMind — Lyria 3 Pro: Create longer tracks in more

- Company: Google DeepMind (deepmind.google)
- Announced: 2026-03-25T16:01:39+00:00
- Category: model-update
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://deepmind.google/blog/lyria-3-pro-create-longer-tracks-in-more/
- Record: https://forck.live/items/1246-lyria-3-pro-create-longer-tracks-in-more
- Subject: Gemini

Google DeepMind introduces Lyria 3 Pro, which enables longer tracks with structural awareness, and expands Lyria to more Google products and surfaces.

### Sakana AI — The AI Scientist: Towards Fully Automated AI Research, Now Published in Nature

- Company: Sakana AI (sakana.ai)
- Announced: 2026-03-25T15:00:00+00:00
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://sakana.ai/ai-scientist-nature
- Record: https://forck.live/items/4691-the-ai-scientist-towards-fully-automated-ai-research-now-published-in-nature
- Subject: Sakana models

Sakana AI announces that its paper on The AI Scientist, which describes an autonomous system capable of performing the entire machine learning research lifecycle, has been published in Nature after a 1.5-year development process.

### Harvey — Take Command of AI Adoption With Harvey's Governance Controls

- Company: Harvey (harvey.ai)
- Announced: 2026-03-25T14:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.harvey.ai/blog/harveys-governance-controls
- Record: https://forck.live/items/17590-take-command-of-ai-adoption-with-harvey-s-governance-controls
- Subject: Harvey

Harvey introduced new admin governance controls and analytics for its AI platform, including updated Usage and Query History APIs, an admin analytics dashboard, and Shared Spaces controls with external request flows, confidentiality settings, and audit logs.

### Harvey — Harvey Raises Growth Round at $11 Billion Valuation Co-led by GIC and Sequoia

- Company: Harvey (harvey.ai)
- Announced: 2026-03-25T12:30:00+00:00
- Category: funding-company-news
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.harvey.ai/blog/harvey-raises-growth-round-at-dollar11-billion-valuation-co-led-by-gic-and-sequoia
- Record: https://forck.live/items/17591-harvey-raises-growth-round-at-11-billion-valuation-co-led-by-gic-and-sequoia
- Subject: Harvey

Harvey announced it raised $200 million at an $11 billion valuation, co-led by GIC and returning investor Sequoia, with participation from existing investors including Andreessen Horowitz, Coatue, Conviction Partners, Elad Gil, Evantic, and Kleiner Perkins. The funding will be used to expand the more than 25,000 custom agents customers run on Harvey and to scale embedded legal engineering teams that help build and optimize those agents inside top law firms and legal departments.

### OpenAI — Inside our approach to the Model Spec

- Company: OpenAI (openai.com)
- Announced: 2026-03-25T10:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://openai.com/index/our-approach-to-the-model-spec
- Record: https://forck.live/items/219-inside-our-approach-to-the-model-spec
- Subject: GPT / ChatGPT / API

OpenAI explains its Model Spec, a public framework for model behavior that balances safety, user freedom, and accountability as AI systems advance.

### Google Research — Vibe Coding XR: Accelerating AI + XR prototyping with XR Blocks and Gemini

- Company: Google Research (research.google)
- Announced: 2026-03-25T09:14:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://research.google/blog/vibe-coding-xr-accelerating-ai-xr-prototyping-with-xr-blocks-and-gemini/
- Record: https://forck.live/items/1353-vibe-coding-xr-accelerating-ai-xr-prototyping-with-xr-blocks-and-gemini
- Subject: Research

Google Research announces Vibe Coding XR, a workflow that combines Gemini with the open-source XR Blocks framework to let users create interactive WebXR applications from natural language prompts in under 60 seconds.

### Vercel — Unified reporting for all AI Gateway usage

- Company: Vercel (vercel.com)
- Announced: 2026-03-25T07:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://vercel.com/blog/unified-reporting-for-your-ai-spend
- Record: https://forck.live/items/20725-unified-reporting-for-all-ai-gateway-usage
- Subject: v0 / AI SDK / AI Gateway

Vercel launched a beta Custom Reporting API for its AI Gateway, available to Pro and Enterprise plan users. The API provides programmatic access to cost, token usage, and request volume across all AI Gateway traffic, including BYOK requests, and allows breakdowns by model, provider, user ID, custom tag, or credential type. One AI platform using the private beta replaced a third-party proxy and saved over $80K.

### Cursor — Self-hosted Cloud Agents

- Company: Cursor (cursor.com)
- Announced: 2026-03-25
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://cursor.com/changelog/03-25-26
- Record: https://forck.live/items/2624-self-hosted-cloud-agents
- Subject: Cursor

Cursor now supports self-hosted cloud agents that keep code and tool execution within the user's own network.

### OpenAI — Introducing the OpenAI Safety Bug Bounty program

- Company: OpenAI (openai.com)
- Announced: 2026-03-25
- Category: safety-policy-update
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://openai.com/index/safety-bug-bounty
- Record: https://forck.live/items/220-introducing-the-openai-safety-bug-bounty-program
- Subject: GPT / ChatGPT / API

OpenAI has launched a Safety Bug Bounty program aimed at identifying AI abuse and safety risks such as agentic vulnerabilities, prompt injection, and data exfiltration.

### Google Research — TurboQuant: Redefining AI efficiency with extreme compression

- Company: Google Research (research.google)
- Announced: 2026-03-24T19:54:00+00:00
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://research.google/blog/turboquant-redefining-ai-efficiency-with-extreme-compression/
- Record: https://forck.live/items/1354-turboquant-redefining-ai-efficiency-with-extreme-compression
- Subject: Research

Google Research introduces TurboQuant, a compression algorithm for large language models and vector search engines, achieving massive compression with zero accuracy loss, along with the QJL and PolarQuant methods.

### Google Research — Mapping the modern world: How S2Vec learns the language of our cities

- Company: Google Research (research.google)
- Announced: 2026-03-24T17:42:00+00:00
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://research.google/blog/mapping-the-modern-world-how-s2vec-learns-the-language-of-our-cities/
- Record: https://forck.live/items/1355-mapping-the-modern-world-how-s2vec-learns-the-language-of-our-cities
- Subject: Research

Google Research introduces S2Vec, a self-supervised framework that transforms complex geospatial data into general-purpose embeddings to predict socioeconomic and environmental patterns across the globe.

### Lovable — A founder's guide to Lovable security

- Company: Lovable (lovable.dev)
- Announced: 2026-03-24T16:44:54+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://lovable.dev/blog/a-founders-guide-to-lovable-security
- Record: https://forck.live/items/12093-a-founder-s-guide-to-lovable-security
- Subject: Lovable / AI app builder

Lovable published a guide for founders on how technical due diligence reviews evaluate applications built with AI tools. The guide outlines five areas assessed in security reviews—data access controls, infrastructure security, vulnerability management, dependency hygiene, and incident readiness—and explains that AI-generated code introduces the same vulnerability categories as human-written code but at higher velocity. It describes Lovable's platform security features including multi-cloud architecture, full tenant isolation with separate Supabase projects per application, and encryption protections.

### fal — Inworld TTS-1.5 Max Now Available on fal

- Company: fal (fal.ai)
- Announced: 2026-03-24T16:00:51+00:00
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://blog.fal.ai/inworld-tts-1-5-max-now-available-on-fal/
- Record: https://forck.live/items/8000-inworld-tts-1-5-max-now-available-on-fal
- Subject: fal platform

fal announced the addition of Inworld TTS-1.5 Max, a text-to-speech model offering low-latency speech generation (time-to-first-audio under ~250ms), improved expressiveness, and support for 15 languages, with pricing at approximately $0.01 per minute.

### Harvey — Building an Agent for Complex Document Drafting and Editing

- Company: Harvey (harvey.ai)
- Announced: 2026-03-24T15:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.harvey.ai/blog/building-an-agent-for-complex-document-drafting-and-editing
- Record: https://forck.live/items/17592-building-an-agent-for-complex-document-drafting-and-editing
- Subject: Harvey

Harvey redesigned its document editing agent to iteratively build a Word document in memory using a proprietary OOXML library, replacing a one-shot parallel sub-agent architecture. The new system gives the model tools to compose edits step by step, supports parallel sub-agents with auto-merge, and includes self-review via a diff tool. It ships in the Harvey for Word Add-In and the web Assistant.

### Lovable — Lovable partners with Aikido to offer affordable pentesting for builders

- Company: Lovable (lovable.dev)
- Announced: 2026-03-24T11:50:55+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://lovable.dev/blog/announcing-pentesting
- Record: https://forck.live/items/12094-lovable-partners-with-aikido-to-offer-affordable-pentesting-for-builders
- Subject: Lovable / AI app builder

Lovable has partnered with Aikido to offer automated penetration testing directly within its platform. The pentest, which previously cost thousands of dollars and took weeks, now takes one to four hours and costs $100 per test. Findings sync into Lovable as actionable issues with AI-generated remediation recommendations, and users can generate an audit-ready report for SOC 2, ISO 27001, and client security questionnaires.

### OpenAI — Helping developers build safer AI experiences for teens

- Company: OpenAI (openai.com)
- Announced: 2026-03-24T11:00:00+00:00
- Category: safety-policy-update
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://openai.com/index/teen-safety-policies-gpt-oss-safeguard
- Record: https://forck.live/items/221-helping-developers-build-safer-ai-experiences-for-teens
- Subject: GPT / ChatGPT / API

OpenAI released prompt-based teen safety policies for developers using gpt-oss-safeguard to moderate age-specific risks in AI systems.

### OpenAI — Powering product discovery in ChatGPT

- Company: OpenAI (openai.com)
- Announced: 2026-03-24T09:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://openai.com/index/powering-product-discovery-in-chatgpt
- Record: https://forck.live/items/223-powering-product-discovery-in-chatgpt
- Subject: GPT / ChatGPT / API

ChatGPT introduces a new shopping feature with richer visuals, powered by the Agentic Commerce Protocol, enabling product discovery, side-by-side comparisons, and merchant integration.

### OpenAI — Update on the OpenAI Foundation

- Company: OpenAI (openai.com)
- Announced: 2026-03-24T09:00:00+00:00
- Category: funding-company-news
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://openai.com/index/update-on-the-openai-foundation
- Record: https://forck.live/items/222-update-on-the-openai-foundation
- Subject: GPT / ChatGPT / API

OpenAI Foundation announces a plan to invest at least $1 billion in curing diseases, economic opportunity, AI resilience, and community programs.

### Hugging Face — A New Framework for Evaluating Voice Agents (EVA)

- Company: Hugging Face (huggingface.co)
- Announced: 2026-03-24T02:01:52+00:00
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://huggingface.co/blog/ServiceNow-AI/eva
- Record: https://forck.live/items/1531-a-new-framework-for-evaluating-voice-agents-eva
- Subject: Platform

ServiceNow AI introduces EVA, an end-to-end evaluation framework for conversational voice agents that jointly scores task accuracy (EVA-A) and conversational experience (EVA-X). They release an initial airline dataset of 50 scenarios and provide benchmark results for 20 systems, finding a consistent Accuracy-Experience tradeoff.

### Inception — Mercury 2 on PinchBench: Fast Diffusion Models and the Personal Agent Era

- Company: Inception (inceptionlabs.ai)
- Announced: 2026-03-24
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://www.inceptionlabs.ai/blog/mercury-2-on-pinchbench
- Record: https://forck.live/items/18523-mercury-2-on-pinchbench-fast-diffusion-models-and-the-personal-agent-era
- Subject: Mercury

Inception evaluated Mercury 2 on the PinchBench benchmark, reporting a 78% task success rate, the fastest execution time in its performance class, and pricing under $1 per million tokens ($0.25 input, $0.75 output). The company states the model's speed comes from parallel refinement rather than token-by-token generation.

### Gradium — Time to First Audio: Measuring and Reducing TTS Latency in Voice Agents

- Company: Gradium (gradium.ai)
- Announced: 2026-03-24
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://gradium.ai/blog/time-to-first-audio
- Record: https://forck.live/items/18325-time-to-first-audio-measuring-and-reducing-tts-latency-in-voice-agents
- Subject: Gradium TTS / Phonon

Gradium published a guide on measuring Time to First Audio (TTFA) for TTS latency in voice agents, including a benchmark comparing its own models to ElevenLabs, Mistral, and OpenAI. The post explains how to accurately measure TTFA by parsing past container metadata and controlling variables such as input text, output format, geographic region, network latency, and warm state. Gradium reports that its models deliver first audio faster than the alternatives while maintaining comparable voice quality.

### LG AI Research — [AAAI 2026] A Design Guide for Organizations Implementing Agentic AI

- Company: LG AI Research (lgresearch.ai)
- Announced: 2026-03-24
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.lgresearch.ai/blog/view?seq=637
- Record: https://forck.live/items/4754-aaai-2026-a-design-guide-for-organizations-implementing-agentic-ai
- Subject: EXAONE

This is a blog post by LG AI Research summarizing a demo track at AAAI 2026 about designing observable and reliable AI agents. It describes three tools: AgentGraph (a debugging interface), AgentSeer (action-level red teaming), and Omega (converting agent choices into natural language explanations). The post also discusses a future roadmap for agent ecosystems, emphasizing transparency, explainability, and trustworthiness. No new AI product or model is announced.

### Replit — Live from Replit HQ Part 2

- Company: Replit (replit.com)
- Announced: 2026-03-23T15:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://replit.com/blog/live-from-hq-agent4-launch-pt2
- Record: https://forck.live/items/17913-live-from-replit-hq-part-2
- Subject: Replit Agent

Replit's Agent 4 launch includes a parallel task system that resolves merge conflicts automatically 90% of the time, an Infinite Canvas for design-to-code prototyping, real-time collaboration visibility, and instant branching using micro VMs. The platform does not charge per seat for collaborators. One company saved over $1 million annually using Replit to automate marketing tasks.

### Vercel — SERHANT.'s playbook for rapid AI iteration

- Company: Vercel (vercel.com)
- Announced: 2026-03-23T04:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://vercel.com/blog/serhants-playbook-for-rapid-ai-iteration
- Record: https://forck.live/items/20727-serhant-s-playbook-for-rapid-ai-iteration
- Subject: v0 / AI SDK / AI Gateway

SERHANT.'s S.MPLE AI product, built on Next.js and Vercel, orchestrates multiple models (Claude Sonnet, Claude Haiku, OpenAI, Gemini) by task to optimize cost and output. The team scaled from a 200-agent pilot to over 900 users without infrastructure changes, using Vercel's AI SDK and AI Gateway for abstraction and unified visibility. The company is now building toward conversational AI agents that combine multiple tasks in a single request.

### Vercel — new.website joins forces with v0

- Company: Vercel (vercel.com)
- Announced: 2026-03-23T04:00:00+00:00
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://vercel.com/blog/new-website-joins-forces-with-v0
- Record: https://forck.live/items/20726-new-website-joins-forces-with-v0
- Subject: v0 / AI SDK / AI Gateway

Vercel announced that new.website, a startup focused on built-in website primitives like forms and SEO, is joining the v0 team. The new.website team will work on making features such as forms, content, and data native and agent-aware within v0.

### Mistral AI — Speaking of Voxtral

- Company: Mistral AI (mistral.ai)
- Announced: 2026-03-23
- Category: open-weight-release
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://mistral.ai/news/voxtral-tts
- Record: https://forck.live/items/3287-speaking-of-voxtral
- Subject: Mistral / Le Chat

Mistral AI announces Voxtral TTS, an open-weights text-to-speech model described as frontier, fast, and adaptable, capable of producing lifelike speech for voice agents.

### OpenAI — Creating with Sora Safely

- Company: OpenAI (openai.com)
- Announced: 2026-03-23
- Category: safety-policy-update
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://openai.com/index/creating-with-sora-safely
- Record: https://forck.live/items/224-creating-with-sora-safely
- Subject: GPT / ChatGPT / API

OpenAI discusses safety measures in building Sora 2 and the Sora app, anchored in concrete protections.