Top249
Posts
Top249: Salesforce launches Agentforce 3 | Models teach themselves | Music industry's AI hunt

Top249: Salesforce launches Agentforce 3 | Models teach themselves | Music industry's AI hunt

Learn more, scroll less. Curated AI/Tech/Business news and articles.

June 24, 2025 • Estimated Reading Time: 4 minutes

24th June 2025

Researchers at MIT introduced SEAL, a new system enabling AI models to generate their own training data and learning instructions, allowing continuous adaptation post-deployment.
The framework empowers AI to self-improve using fresh data, reducing reliance on static pre-training and potentially enabling more robust long-lived models.
Though still experimental, SEAL signals a shift toward AI that can learn on the fly, adapting to new environments without full retraining.

Salesforce released Agentforce 3, featuring a new “Command Center” for real-time observability and control over AI agent operations.
It supports emerging interoperability standards like Protocols for Multi-Component workflows and Agent-to-Agent communication, enabling secure, low-code integrations across enterprise tools.
The update aims to help organizations scale digital labor confidently, with improved visibility, governance, and support for monitoring partners.

Andy Konwinski, co-founder of Databricks and Perplexity, pledged $100 million of personal funds to launch a new AI research institute named Laude.
The institute will support early-stage, high-impact projects and is already backing Ion Stoica’s new lab, signaling deep commitment to advancing foundational AI research.
This investment reflects growing private-sector funding aimed at driving innovation in AI infrastructure and capabilities.

In response to rising AI-generated tracks, the music industry is deploying tools like Vermillio’s TraceID to identify, label, license, or restrict AI-created music.
Platforms including Deezer, YouTube, and SoundCloud are implementing metadata tagging and detection systems, as AI content begins to account for roughly 20% of daily uploads.
Beyond takedowns, these systems aim for transparent licensing and proactive management of AI-generated content—tackling issues at distribution and training levels.

Microsoft introduced a new evaluation methodology that predicts and explains model performance on unfamiliar tasks by generating interpretable ability profiles.
Podcast insights highlight the use of automated AI testing, red-teaming, and layered defense systems that combine human oversight with AI-generated assessments.
These approaches reflect a push to improve model safety and reliability by bridging scientific rigor with industrial deployment practices.

Microsoft proposed integrating lessons from fields like biology and aerospace to improve AI evaluation methods.
By adopting domain-specific benchmarks and multi-metric analysis, researchers aim to move beyond single-number performance metrics.
This approach encourages safer, more robust AI development rooted in continuous learning and cross-disciplinary rigor.

China launched a nationally coordinated initiative in 2009 called “Ten Cities, Thousand Vehicles,” expanding it to 25 cities to jumpstart EV adoption in public transport and incentivize manufacturers.
Massive state investment followed: over $28 billion in subsidies and tax breaks fueled both production and consumer uptake, while local regulations (like green license plates and vehicle quotas) accelerated EV use.
Combined with rapid build-out of charging infrastructure and dominance in battery supply chains, China now leads globally—producing nearly 60% of EVs and 80% of batteries—turning EVs into a mass-market norm.

Till next time.