arXiv:2606.24963v1 Announce Type: new Abstract: Fine-tuning Multimodal Large Language Models (MLLMs) on specialized tasks often leads to catastrophic forgetting of their general capabilities. Existing model merging methods to combat this are often heuristic or use sub-optimal objectives. We propose CurvatureGuided Mixing (CGM), a theoretically grounded framework that merges pre-trained and fine-tuned models. CGM formulates a joint optimization objective and uses a second-order (Hessian) approximation of the loss landscapes to analytically derive an optimal, closed-form "soft mixing" ratio. This ratio intelligently blends parameters based on their relative task-specific curvatures. We also introduce CGM$\dagger$, a robust "hard mixing" variant that performs sparse parameter selection guided by a novel, curvature-aware score. Experiments on LLaVA-1.5 and Qwen2.5VL across multiple downstream tasks show that CGM and CGM$\dagger$ consistently improve the trade-off between task specialization and general knowledge retention over existing methods. Code is available at github.com/zzsyjl/CGM-ECCV-2026.

arXiv:2606.25084v1 Announce Type: new Abstract: Multimodal Large Language Models (MLLMs) have redefined visual understanding by combining vision encoders with large-scale language models. This unified architecture enables strong performance on tasks like image captioning, visual question answering, and multimodal dialogue, often in zero- and few-shot settings. Their general-purpose capabilities and flexible interfaces make MLLMs a promising foundation for real-world vision-language applications. Assistive AI aims to help users interact with their environments through natural language. These scenarios demand robust visual recognition, contextual reasoning, and multilingual comprehension-capabilities that MLLMs are believed to offer. However, their effectiveness in assistive settings remains to be fully understood. In this work, we explore whether MLLMs can support Assistive AI by evaluating state-of-the-art models on real-world tasks: recognizing everyday objects like currency, answering questions based on scene text, and reading visually presented content across multiple languages. To this end, we developed a system, NetraLink, using a head-mounted GoPro to capture real-world egocentric data, and collected a benchmark covering these assistive scenarios. Our findings provide a comprehensive diagnostic of current MLLMs, highlighting their strengths and limitations in enabling assistive technologies grounded in visual perception and language interaction.

A new OpenAI research paper shows how AI agents are transforming work, enabling longer, more complex tasks and expanding productivity across roles.

Discussion | Link

Micah Lasher won New York's 12th congressional district primary race with 39.1% of the vote. Bloomberg/Getty Images Micah Lasher won New York's 12th congressional district primary race among eight candidates. Tech giants spent millions to influence the Manhattan race and sway AI regulation outcomes. The Manhattan congressional seat became an AI-regulation showdown with two candidates at its center. The new Democratic nominee for a congressional seat in Manhattan had a strong message for the AI industry on Tuesday night. After defeating a crowded field of eight candidates in New York's 12th congressional district, Micah Lasher, who won by 39.1%, took aim at the tech giants and their allies who spent heavily to shape the outcome of the deep-blue seat. "I have some news for the two big AI companies who've taken such an unusual interest in who won this congressional seat," Lasher said at a rally at Jacob's Pickles, an NYC staple famous for elevating the pickle from a garnish to a main course, following his victory. "I won't be taking my cues from either of you when it comes to protecting our kids, our jobs." The swipe at AI companies didn't come out of nowhere. On the trail, Lasher cast himself as a skeptic of Silicon Valley's push for lighter-touch regulation, stating on his website that AI could "displace workers, exacerbate inequalities, and pose a threat to our environment and public safety." He has also raised concerns about the rapid expansion of AI data centers and the industry's growing energy demands. The race in Manhattan became an unlikely battleground in the fight over how Washington should regulate AI, illustrating a schism in Silicon Valley. Millions of dollars flowed into the contest, much of it aimed at either helping or hurting Assemblymember Alex Bores, a Democrat and former Palantir employee who backed stronger AI safeguards, as competing factions of the tech world backed different candidates and different visions for AI policy. Bores quit Palantir during Donald Trump's first term, citing concerns about the company's work on immigration enforcement. According to Federal Election Commission filings, Think Big, a super PAC opposed to additional AI regulations backed in part by leaders at OpenAI and Andreessen Horowitz, spent about $8 million to prevent Bores from winning. Meanwhile, tech giants backing more AI safety regulations, including the Jobs and Democracy PAC, supported by donors with ties to Anthropic and Adobe, spent more than $13 million to boost Bores' candidacy. The clash came as New York emerges as one of the more aggressive states in the country in regulating the AI industry, and as data centers have become a growing flash point. New York State lawmakers have advanced proposals that would temporarily halt the issuance of permits for large new data centers while officials study their impact on the electric grid, utility bills, water consumption, and climate goals. Lasher is a cosponsor of New York's Responsible AI Safety and Education Act, a proposal aimed at placing safeguards on advanced AI systems, which is the same legislation that made Bores a target for some AI industry groups. Anthropic, OpenAI, and the Lasher campaign did not immediately respond to requests for comment. Read the original article on Business Insider

Jagmeet Singh / TechCrunch: Hang Ten Systems, led by former Infosys CEO Vishal Sikka, launches with a $32M seed led by Mayfield to help enterprises use AI to run software at a lower cost  —  For decades, IT services firms made billions of dollars by allowing companies to outsource tech tasks like customizing, integrating, and maintaining enterprise software.

As ASML CEO Christophe Fouquet told TechCrunch in May, what China can currently buy are older-generation deep ultraviolet tools — gear first shipped about a decade ago — the same machines the MATCH Act would now put off-limits.

<p><strong><a href="https://github.com/simonw/browser-compat-db">simonw/browser-compat-db</a></strong></p> Inspired by Mozilla's <a href="https://developer.mozilla.org/en-US/blog/introducing-mdn-mcp-server/">new MDN MCP service</a> - <a href="https://github.com/mdn/mcp">source code here</a> - I decided to try converting their comprehensive <a href="https://github.com/mdn/browser-compat-data">mdn/browser-compat-data</a> repository full of browser compatibility data into a SQLite database.</p> <p>This new GitHub repo includes a Claude Code for web (Opus 4.8) <a href="https://github.com/simonw/browser-compat-db/blob/main/build_db.py">generated script</a> for doing that using <a href="https://github.com/simonw/sqlite-utils">sqlite-utils</a>.</p> <p>I wanted the resulting ~66MB SQLite database to be available via the GitHub CDN with open CORS headers. GitHub releases don't have those, but any file stored in a regular GitHub repository does - so I had Codex Desktop (GPT-5.5) build <a href="https://github.com/simonw/browser-compat-db/blob/main/.github/workflows/build-db.yml">a GitHub Actions workflow</a> that builds the database and then force-pushes it to a <code>db</code> "orphan" branch.</p> <p>You can download the resulting database <a href="https://github.com/simonw/browser-compat-db/blob/db/browser-compat.db">from here</a>, and since it's hosted with open CORS headers you can also <a href="https://lite.datasette.io/?url=https://github.com/simonw/browser-compat-db/blob/db/browser-compat.db#/browser-compat/releases_tree">explore it with Datasette Lite</a>. <p>Tags: <a href="https://simonwillison.net/tags/github">github</a>, <a href="https://simonwillison.net/tags/mozilla">mozilla</a>, <a href="https://simonwillison.net/tags/projects">projects</a>, <a href="https://simonwillison.net/tags/github-actions">github-actions</a>, <a href="https://simonwillison.net/tags/datasette-lite">datasette-lite</a>, <a href="https://simonwillison.net/tags/ai-assisted-programming">ai-assisted-programming</a>, <a href="https://simonwillison.net/tags/model-context-protocol">model-context-protocol</a>, <a href="https://simonwillison.net/tags/mdn">mdn</a></p>

Brock E.W. Turner / Axios: Trase, which is building an operating system and infrastructure layer for AI agents in industries like health care and defense, raised a $107M seed led by Arch  —  Trase, an AI operating system company, has raised $107 million in seed funding led by Arch Venture Partners, CEO Grant Verstandig tells Axios Pro exclusively.

Discussion | Link

Backed by Mayfield and Aramco Ventures, Vishal Sikka’s new venture brings together veterans from SAP, Infosys, and VianAI.

Tim Bradshaw / Financial Times: Paris-based digital health insurance startup Alan raised €480M led by Prosus at a €5.5B valuation, and is set for €1B+ in annual recurring revenue by year's end  —  Paris-based group raises one of Europe's biggest non-AI start-up rounds this year

Right now he's merely a several-hundred-billionaire, according to Bloomberg's Billionaires Index.

Jack Healy / New York Times: Utah Senate President J. Stuart Adams, who championed a huge data center beside the Great Salt Lake, lost his Republican primary on Tuesday after voter backlash  —  The upset of J. Stuart Adams represented one of the strongest examples of how the data center issue has affected elections.

Zoox refreshed its robotaxi and said it's prepared to produce up to 100 vehicles a week. Zoox Zoox revealed a refreshed robotaxi design that the company says will improve rider experience. The update comes as the company prepares to ramp up vehicle production at its Hayward factory. The facility has the capacity to produce 10,000 robotaxis a year. Amazon's Zoox is getting ready to ramp up vehicle production with a newly refreshed robotaxi design as the company seeks to expand its service areas in the US. Zoox unveiled several tweaks to its purpose-built robotaxi on Wednesday, including changes to the interior color, more ergonomic seats, larger cupholders, a more vivid touchscreen, and two-way audio capabilities designed to improve communication with riders and first responders. Zoox said it can produce up to 100 of the newly updated robotaxis a week to support its expansion plans for this year, pending regulatory approval. A spokesperson for Zoox said the regulatory approval refers to a pending petition with the National Highway Traffic Safety Administration. The petition seeks a temporary exemption from some federal safety standards that assume a vehicle has a human driver and traditional driving controls. The core robotaxi design remains unchanged. Zoox said the updates are designed to make rides feel calmer and more intuitive as it prepares to put more vehicles on the road. The lighter interior color scheme is meant to reduce "visual distractions" and make it easier for riders to spot items left behind, such as phones or keys, Zoox said. Zoox updated the interior to include a lighter color scheme, more ergonomic seats, and other adjustments. Zoox "These robotaxis will join the fleet across our markets and become available to riders later this year as they come off the production line," Zoox said. The company said in its announcement that the updated vehicle is its production-intent robotaxi and will be built at its Hayward, California, factory. Business Insider previously reported that Zoox's 220,000-square-foot factory can produce more than 10,000 vehicles a year. A Zoox spokesperson told Business Insider that the company does not need to build 10,000 robotaxis at the moment. "We are ramping production in a deliberate, phased manner to safely meet the strong consumer demand and regulatory requirements," the spokesperson said. Zoox provides free robotaxi rides to the public in limited parts of Las Vegas and San Francisco as the company continues to collect rider feedback. The company said that it plans to expand to Austin and Miami later this year. Read the original article on Business Insider

Meta CEO Mark Zuckerberg. Bloomberg/Getty Images Meta reassigned thousands of engineers to units focused on AI training last month. Now, Meta is letting people leave the unit if they want to. The move comes after CTO Andrew Bosworth addressed morale dips at Meta this month. Meta is walking back its stance on forcing engineers to join a task force focused on AI training, according to an internal memo obtained by Business Insider and four people familiar with the matter. Last month, Meta reassigned 7,000 employees to units such as an Applied AI task force to help train Meta's coming AI models. On Wednesday, Meta sent a memo about this task force, saying the company will now "defer to each individual's choice." The company sent the email to employees who had been "drafted," as some described its Applied AI task force. "As I emphasized before, personal agency will remain at the heart of all opportunities at Meta: we will support employees in whatever decisions they make," the memo said. "Of course, we'd prefer everyone to stay and push to SOTA together, but we defer to each individual's choice," it read, referring to state-of-the-art. The memo went on to say that people in the unit would have preferential placement in other parts of the company due to staffing shortages. Meta declined to comment for this story. Some employees on Blind called the memo an "undraft." The task force faced significant backlash last month from employees who compared the job to data labeling. The reversal comes after chief technology officer Andrew Bosworth addressed a broader morale crisis at the company. During an internal "Tuesdays with Boz" session on June 2, Bosworth told employees that morale was "probably one of the worst it's ever been" in Meta's 20-year history, Business Insider previously reported. In May, Meta laid off 10% of its staff, or 8,000 people. Have a tip? Contact Charles via email at crollet@businessinsider.com or on Signal and WhatsApp at 628-282-2811. Contact Pranav via encrypted messaging app Signal at +1-408-905-9124, or email him at pdixit@businessinsider.com or pranavdixit@protonmail.com. Use a personal email address, a nonwork WiFi network, and a nonwork device; here's our guide to sharing information securely. Read the original article on Business Insider

GTA6 might be an outlier, though—at least for now.

In its first earnings report since going public, the AI chipmaker forecast a narrower gross margin in its core business, scaring investors.

Google’s Search history update stores media uploads from your interactions, like images used in reverse image searches, for training its AI models.

The silicon race is heating up amid the struggle to keep up with demand.