Another AI Newsletter: Week 28
This week’s roundup covers new AI tools, major industry moves, and research breakthroughs driving the next wave of innovation.
Another AI Newsletter - Week of July 05, 2025 to July 12, 2025
Major Product/Tool Releases
Google introduces Flow AI filmmaking tool and Gemini photo-to-video feature
July 2025 | Google MENA Blog
Google announced two generative media products: *Flow*, a new “AI filmmaking tool” for advanced models, and a photo-to-video capability in its Gemini chat app [blog.google]. With the Gemini feature, users can “transform their favorite photos into dynamic eight-second video clips with sound” using the Veo 3 model [blog.google]. Both Flow and the enhanced Gemini model (with Veo 3) are launching first in the Middle East/North Africa region, targeting content creators and artists.
Why it matters: These releases bring cutting-edge video-generation tools to users, enabling anyone to animate images and create short films. By packaging Flow and the photo-to-video feature into accessible apps, Google is pushing AI-powered creative tools toward mass adoption.
Perplexity AI debuts ‘Comet’ AI-powered web browser
July 2025 | Perplexity Blog
Perplexity AI launched Comet on July 9 as a browser built around its AI engine. Announced on its blog, “Today we are launching Comet. Comet is a web browser built for today’s internet” [www.perplexity.ai]. Comet allows subscribed users (Perplexity Max) to ask questions and issue commands via text or voice while browsing, effectively integrating AI search into the browser. It provides natural-language answers and multi-step assistance within web pages, all backed by Perplexity’s large language model.
Why it matters: Comet represents a shift from traditional browsing to conversational search. By embedding an AI assistant directly in the browser, Perplexity aims to rival incumbents (Google, Microsoft) and make web navigation more interactive. This product foreshadows a new class of AI-powered browsers where search, Q&A, and apps blend seamlessly.
Google adds Gemini’s Veo 3 image-to-video generation to Workspace
July 2025 | Google Workspace Updates
On July 11, Google announced that Workspace users can now use Gemini’s Veo 3 AI model to generate short videos from images [workspaceupdates.googleblog.com]. Enterprise customers (Business Standard/Plus, Enterprise, Education, etc.) can upload a still photo and create an 8-second video, complete with realistic audio (birds, traffic, dialogue) and physics-based motion [workspaceupdates.googleblog.com]. The Workspace update explicitly notes the “photo to video feature that allows you to upload a photo and create an 8-second video,” with a daily usage limit (3–5 per day) for business users [workspaceupdates.googleblog.com].
Why it matters: By bringing advanced video-generation to its productivity suite, Google lets organizations effortlessly create illustrative videos for work. This embeds generative AI directly into enterprise apps, enabling new workflows (e.g. auto-generated explainer videos from graphics) and signaling that creative AI is becoming a standard business feature.
Breakthrough Research or Papers
Open Vision Reasoner: Transferring Linguistic Cognitive Behavior for Visual Reasoning
July 2025 | ArXiv
This paper introduces a two-stage training method that first fine-tunes a language model and then applies nearly 1,000 steps of multimodal reinforcement learning to teach the model advanced visual reasoning. The resulting model, Open-Vision-Reasoner (OVR), achieves state-of-the-art results on multiple reasoning benchmarks – for example, it attains 95.3% accuracy on MATH500 and 54.6% on MathVerse when tested on challenging multimodal tasks [lonepatient.top]. These metrics underscore how the approach effectively transfers LLM reasoning skills into vision-language models.
Why it matters: OVR demonstrates a practical way to endow vision-language systems with deep reasoning abilities, moving the needle on multimodal AI. By showing large gains across benchmarks, it points toward more capable agents that can understand and reason about complex visual/math problems using LLM-derived reasoning and reinforcement learning.
Response Attack: Exploiting Contextual Priming to Jailbreak Large Language Models
July 2025 | ArXiv
This work uncovers a new vulnerability in LLM safety: contextual priming. It introduces Response Attack (RA), which feeds a subtly malicious response (generated by a helper LLM) into the dialogue history before a target model is queried. In experiments on eight open and closed-source LLMs, RA “consistently outperforms seven state-of-the-art jailbreak techniques, achieving higher attack success rates”. To counter this, the authors release a context-aware fine-tuning dataset: after safety-tuning with it, models show a dramatic drop in RA’s success rate while maintaining overall performance.
Why it matters: This research reveals a powerful new way that LLMs can be manipulated to produce harmful content (through indirect priming), and simultaneously offers a defense. Identifying such a potent jailbreak technique – and mitigating it – is critical for real-world LLM deployment, highlighting both the risks and remedies in model alignment and security.
MemoryAgentBench: A Benchmark for Evaluating Memory in LLM Agents
July 2025 | ArXiv
The authors address the lack of tests for LLM-based agents’ memory. They “identify four core competencies essential for memory agents: accurate retrieval, test-time learning, long-range understanding, and conflict resolution”, and introduce MemoryAgentBench, the first benchmark to cover all four. This benchmark combines reformulated and newly created datasets to systematically challenge an agent’s long-term memory skills. Initial results are sobering: existing memory-augmented agents – from simple retrieval-augmented models to those with external memory modules – “fall short of mastering all four competencies”.
Why it matters: By providing a comprehensive testbed for memory in LLM agents, this work lays the groundwork for future advances in agent intelligence. MemoryAgentBench equips researchers with a way to measure and push progress on long-term memory in conversational and task-driven AIs – an essential capability for agents that must remember and reason over extended interactions.
Real-World Use Cases and Demos
Neyox.ai Launches Voice AI Agents to Supercharge U.S. Real Estate Sales
July 2025 | PR.com
India-based AI startup Neyox.ai has rolled out “Voice AI Agents” in the U.S. real estate market to automate cold outreach. These agents handle thousands of outbound calls 24/7—qualifying leads, booking appointments, and following up—so brokers can focus on selling (source). “U.S. real estate agents spend a significant portion of their time chasing leads… Our AI Voice Agents eliminate this bottleneck,” says founder Neeraj Parnami.
Why it matters: Automating repetitive customer contact at scale shows how AI can boost efficiency and lead conversion in sales-heavy industries. By freeing agents from routine tasks, the technology demonstrates tangible time savings and growth potential for real estate (and similar sales-driven) businesses.
Chinese Steelmaker Revs Up Efficiency with AI-Powered Visual Inspection
July 2025 | People’s Daily (Xinhua)
Shougang Group, one of China’s largest steel producers, deployed an AI-driven visual inspection system across its production line (en.people.cn). This computer-vision system turns steelmaking “from labor-driven to model-driven,” raising output efficiency by over 20% and cutting defect rates by 35% (en.people.cn). The upgrade, highlighted at China’s 2025 Global Digital Economy Conference, exemplifies the push toward smart manufacturing in heavy industry.
Why it matters: These gains – double-digit productivity lift and far fewer errors – translate directly into higher yields and lower waste for a legacy industry. The case underscores AI’s real-world impact on factory floors, proving that machine vision can dramatically improve quality and throughput in large-scale manufacturing.
NIX Champions Scalable AI Deployment at Tech Conference
July 2025 | PR.com
At a Boston MTLC tech event, software integrator NIX shared how enterprises should adopt AI for real impact. NIX’s AI lead Max Ushchenko emphasized that success comes from “building something that lasts—something that can scale, stay transparent, and support the business as it grows,” not just one-off pilots (www.pr.com). He outlined how NIX helps clients establish strategic AI frameworks and data practices so projects deliver measurable results, rather than fizzling out.
Why it matters: By focusing on durable, enterprise-grade implementations (rather than hype projects), this approach highlights the importance of long-term ROI. It shows how aligning AI initiatives with business processes and data quality is crucial for achieving consistent value and avoiding “pilot purgatory” in real-world deployments.
Agentic AI and Reasoning Advances
Spatio-Temporal LLM: Reasoning about Environments and Actions
July 2025 | ArXiv
This new multimodal model tackles the challenge of understanding both an agent’s full environment and its recent actions together (chatpaper.com). The authors create a large “REA” dataset (Reasoning about Environment and Actions) and introduce a specialized ST-LLM architecture (with spatial and temporal projection modules) that markedly improves performance on tasks requiring that holistic spatio-temporal reasoning. Their experiments show this model significantly outperforms prior approaches on benchmarks demanding whole-scene awareness.
Why it matters: Integrating broad contextual understanding into AI agents is crucial for real-world autonomy; this work brings AI closer to human-like planning by enabling coherent, long-range reasoning about environments and actions.
Open Vision Reasoner: Transferring Cognitive Behavior for Visual Reasoning
July 2025 | ArXiv
Open Vision Reasoner adapts the reasoning strategies of large language models to visual tasks (chatpaper.com). It essentially transfers LLM-style cognitive processes (like “chain-of-thought” reasoning) into the image domain, enabling the model to perform multi-step inference directly on visual inputs. This approach substantially improves accuracy on challenging vision-and-language benchmarks by allowing the system to “think out loud” about images as it would about text.
Why it matters: By bringing proven LLM reasoning into vision systems, this advance lets visual AI agents tackle complex tasks that require logical, stepwise analysis of an image, closing a gap between language-driven intelligence and visual understanding.
DreamVLA: A Vision-Language-Action Model Dreamed with Comprehensive World Knowledge
July 2025 | ArXiv
DreamVLA builds a unified vision–language–action agent equipped with an internal “dreamed” world model (chatpaper.com). It uses imagined future trajectories (simulated world knowledge) during training to endow the agent with a rich understanding of how visual observations, language cues, and possible actions interrelate. Initial results demonstrate that DreamVLA greatly outperforms baseline agents on long-horizon tasks requiring integrated perception, language, and planning.
Why it matters: Giving an AI agent a built-in predictive model of the world dramatically enhances its autonomous decision-making. DreamVLA’s world-informed planning represents a big step toward truly self-directed AI that can reason ahead and act independently in complex, multimodal environments.
Thought Leadership and Commentary
Are AI existential risks real—and what should we do about them?
July 2025 | Brookings
Brookings fellow Mark MacCarthy examines long-standing fears that future AI could threaten humanity. He reminds readers of warnings from figures like Stephen Hawking that superintelligent AI might eventually “outsmart financial markets… out-invent human researchers… [and] develop weapons we cannot even understand” (www.brookings.edu). MacCarthy argues policymakers often dismiss these scenarios as overblown, but he echoes experts’ caution that the long-term future “depends on whether [AI] can be controlled at all” (www.brookings.edu).
Why it matters: This analysis injects urgency into AI policy debates by highlighting existential risk. It underscores the need for proactive safety measures and governance, since unchecked AI could dramatically reshape society if we fail to manage its power.
Using AI to advance skills-first hiring
July 2025 | Brookings
Brookings scholars Papia Debroy and Byron Auguste argue that AI should be used to make hiring fairer, not replace workers. They propose a “skills-first” approach, noting that today’s workforce includes 70+ million “STARs” (Skilled Through Alternative Routes) who lack traditional credentials. In their vision, AI acts as “amplified intention”: if applied “equitably and transparently,” it can shift hiring “from exclusion to inclusion” (www.brookings.edu). For example, AI could analyze millions of job postings and employee records to “refine skill taxonomies” and help employers create new career paths, thus “broadening access to jobs that require valuable but often overlooked skills” (www.brookings.edu).
Why it matters: This perspective highlights AI’s potential to democratize opportunity. Instead of automating people out of jobs, AI tools could match workers to roles based on real abilities, expanding mobility for under-leveraged talent and reshaping the labor market to be more inclusive.
Reimagining economic progress in the age of AI
July 2025 | OpenAI Forum
An OpenAI forum event recap explores how we measure prosperity in an AI world. Economist Erik Brynjolfsson introduces “GDP-B,” a new metric to capture the value of free AI services. As he puts it, our current GDP “measures many things, but not services with zero price… AI assistants. GDP-B quantifies this previously invisible consumer surplus” (forum.openai.com). The article suggests adopting such innovative measures and embracing human-AI complementarity. It concludes that by using metrics like GDP-B and fostering entrepreneurial “dynamism,” AI can be guided to “serve as a force for collective benefit” (forum.openai.com).
Why it matters: This analysis challenges conventional economic thinking. By accounting for the billions of dollars of value we get from free AI tools, leaders can better understand AI’s real impact on society. The ideas here point toward policies that focus not just on raw output but on broad social welfare, ensuring that AI-driven innovation lifts everyone.
AI Safety and Ethics Developments
International AI Safety Report 2025 Released
July 2025 | Gov.UK
The UK’s Department for Science published the International AI Safety Report 2025, prepared by 96 global experts. It warns of diverse AI risks (e.g. cybersecurity threats, disinformation, economic disruption) and provides science-backed safeguards. The report describes itself as an “unprecedented effort to establish an internationally shared scientific understanding of risks from advanced AI and methods for managing them” (www.gov.uk).
Why it matters: This first-of-its-kind global review codifies expert consensus on AI dangers and lays out concrete safety and governance measures, guiding policymaking ahead of major AI summits.
EU Unveils Voluntary AI Code of Practice
July 2025 | Reuters
The European Commission announced a voluntary code of practice to help firms comply with the EU AI Act. The code “focuses on transparency, copyright, safety, and security,” and is aimed at major AI providers (Google, Meta, OpenAI, etc.) (www.reuters.com). Although not mandatory, signatories gain a safe harbor, and the code will later be phased in as binding standards for general-purpose AI providers.
Why it matters: By setting unified compliance norms and encouraging industry buy-in, the EU code helps operationalize ethical AI rules (e.g. curbing harmful use and promoting model transparency) and paves the way for global AI safety standards.
AI Now Institute’s “Artificial Power” Report Calls for Oversight
July 2025 | AI Now Institute
The nonprofit AI Now Institute published a 2025 “Landscape Report” titled Artificial Power that stresses community action. It observes that the post–“AI boom” rush to embed AI everywhere has given tech companies and “the tech oligarchs” in Silicon Valley “power that goes far beyond their deep pockets,” and it “offers concrete strategies for community organizers, policymakers, and the public to change this trajectory” (ainowinstitute.org). The report includes analyses of bias, surveillance, and calls for democratic control of AI systems.
Why it matters: By highlighting how unchecked AI development amplifies power imbalances and urging citizen-driven solutions, the report underlines the need for transparency, accountability, and grassroots engagement in AI governance.
Industry Investment and Business Moves
Google hires Windsurf execs in $2.4B deal to advance AI coding ambitions
July 2025 | Reuters
Alphabet’s Google struck a surprise deal with AI coding startup Windsurf: it is paying $2.4 billion in license fees for Windsurf’s AI code-generation technology while hiring key Windsurf talent into Google DeepMind (www.investing.com). Windsurf CEO Varun Mohan, co-founder Douglas Chen, and members of their R&D team are joining Google’s AI lab, while Windsurf itself remains independent under this non-exclusive technology licensing arrangement (www.investing.com). This deal follows failed acquisition talks between Windsurf and OpenAI (which valued Windsurf at ~$3B) (www.brecorder.com), underscoring the intense competition and high stakes in AI-powered software development tools.
Why it matters: Google’s move shows how Big Tech is aggressively securing cutting-edge AI assets. By licensing technology and “acqui-hiring” top AI talent, Google accelerates its Gemini coding initiatives without a full buyout. The deal highlights a broader trend: tech giants are structuring strategic partnerships and talent acquisitions to quickly absorb key AI capabilities, positioning themselves in the fast-growing AI code-generation market (www.brecorder.com).
Amazon mulls multi-$B investment to deepen AI partnership with Anthropic
July 2025 | Business Standard
E-commerce giant Amazon is reportedly considering another multi-billion-dollar investment in AI startup Anthropic (www.business-standard.com). The Financial Times (via media reports) says this potential funding round would build on the ~$8 billion Amazon has already committed to Anthropic (roughly $4B in 2023 and another $4B in 2024) (www.business-standard.com). By expanding its stake, Amazon aims to keep its position as one of Anthropic’s top backers (surpassing Google’s $3B+ investment). This move reinforces the strategic AWS–Anthropic alliance, which includes Amazon supplying custom AI chips and supporting Anthropic’s model development.
Why it matters: The news highlights the escalating AI funding war among tech titans. Additional Amazon funding would cement a deep partnership in AI model development, ensuring Amazon’s influence over Anthropic’s next-generation systems. It signals that cloud and retail leaders like Amazon are willing to invest tens of billions in AI R&D – an arms race to lock in leading AI innovations and infrastructure ahead of competitors.
Elon Musk’s xAI eyes up to $200B valuation in new fundraise
July 2025 | Reuters
Media reports, citing the Financial Times, indicate Elon Musk’s AI startup xAI is preparing a major funding round that could value the company at $170–200 billion (www.reuters.com) – roughly ten times the valuation from mid-2024. Saudi Arabia’s sovereign fund is expected to play a significant role in the planned raise, although Musk tweeted that xAI “has sufficient capital” at present (www.reuters.com). xAI has already secured about $5B in debt financing and $5B in equity for expanding AI data centers in recent months. It projects around $1B in revenue by the end of 2025, with an $18B-plus spend planned for AI infrastructure (www.reuters.com).
Why it matters: A ~$200B valuation for a private AI startup would dwarf almost all previous funding rounds, underlining the massive amounts of capital flooding the AI sector. It underscores Musk’s bet on AI as the next era-defining technology and intensifies the high-stakes competition among AI labs. The news (and Musk’s acquisition of Twitter/X) suggests xAI is positioning itself as a major ChatGPT rival. If consummated, this mega-round would reshape the finance side of AI – signaling that governments and investors are ready to bankroll extraordinarily large AI ventures to capture the future of the industry (www.reuters.com) (www.reuters.com).
Regulatory & Policy
EU launches voluntary AI code of practice to aid compliance
July 2025 | Reuters
The European Commission introduced a voluntary Code of Practice to help firms align with the bloc’s AI Act, with guidance emphasizing transparency, copyright, safety and security [www.reuters.com]. Participation is optional, but as Reuters notes “only signatories will benefit from legal certainty” under the framework [www.reuters.com].
Why it matters: The code provides companies a clear compliance pathway ahead of the AI Act’s August 2025 enforcement, effectively tying major AI developers to the EU’s strict rules and promoting global best practices for safe AI deployment.
Czech Republic bans Chinese AI tool DeepSeek in public sector
July 2025 | Reuters
Prime Minister Petr Fiala announced that Czech government agencies are barred from using DeepSeek, a Chinese AI service, citing “concerns over data security” [www.reuters.com]. The government warned that as a Chinese firm DeepSeek “is required to cooperate with Chinese government authorities,” potentially exposing “sensitive data” to Beijing [www.reuters.com]. Its privacy policy shows it “collects and stores user data ... on computers in China,” raising “substantial security and privacy issues” for Western users [www.reuters.com].
Why it matters: This echoes similar bans in Germany, Italy and the Netherlands, reflecting growing Europe-wide caution toward Chinese AI. The decision underscores how data sovereignty and national security concerns are shaping AI policy.
UN report urges global measures against AI deepfakes
July 2025 | Reuters
A United Nations (ITU) report presented at the “AI for Good” summit calls for “stronger global measures to detect and counter deepfake content” (www.reuters.com). It warns that AI-generated images, video and audio could influence elections or enable fraud, and specifically urges “implementing digital verification tools” and robust content authentication standards to verify authenticity (www.reuters.com). Experts stress that international collaboration (e.g. watermarking and provenance tracking) is needed as AI-enabled misinformation spreads.
Why it matters: By demanding new verification standards, this report signals an urgent global response to the deepfake epidemic. Ensuring media authenticity will be critical to protecting elections, public trust and democracy from malicious AI-driven fake content.
Powered by OpenAI Deep Research API