DEV Community: Jaromir Tesar

Claude’s meeting presence turns Anthropic’s office strategy into execution

Jaromir Tesar — Thu, 12 Nov 2026 15:03:01 +0000

Quick take

Anthropic’s broader expansion includes Claude Sheets, Claude Whiteboards and Claude Data, which create spreadsheets, visual workspaces and dashboards.
Anthropic product lead Nicholas Lin said Claude Meetings was tested internally for about two months and joined nearly all of Anthropic’s meetings.
Anthropic says Claude identifies itself, posts a notice in meeting chat, requires administrator authorization for external guests and leaves if one enters.
Anthropic says Claude can connect with Adobe, Atlassian and HubSpot, potentially linking meeting discussions to business documents and operational systems.

Why it matters

For enterprise teams, Claude is moving from a tool that records meetings to one that helps determine and build what happens next across workplace systems. The next test is whether employees trust its generated artifacts and whether companies can control access, retention and approval so tentative comments do not become commitments.

Anthropic is pushing Claude beyond meeting summaries and into the work itself, giving the assistant a role in converting workplace conversations into plans, documents and assigned tasks while participants are still talking.

TechCrunch reported that Anthropic has launched Claude Meetings in beta for supported Google Meet calls. Teams on Anthropic’s Team and Enterprise plans can add Claude as a visible participant. The assistant follows the conversation, drafts plans, builds timelines on a collaborative whiteboard, answers questions in meeting chat and retrieves information from connected workplace applications.

1515 Third Street · Coolcaesar · via wikipedia · CC BY 4.0

The strategic importance is not the presence of another AI notetaker. It is the attempt to shorten the distance between a decision and the operational work required to act on it. If Claude can turn a discussion into a plan, timeline, spreadsheet or dashboard before the meeting ends, Anthropic gains a stronger position in the software layer where companies organize and execute work.

From records to workplace output

Anthropic’s product announcement presents Claude Meetings as part of a broader expansion that includes Claude Sheets, Claude Whiteboards and Claude Data. The company describes Claude as an expert colleague that can connect information across documents, meetings and decisions, a positioning also emphasized on Anthropic’s launch-event page.

That positioning matters because the established market for meeting assistants has largely focused on recording, transcription and recap generation. Those features can save time, but they often leave employees with a second task: moving the useful parts of a conversation into project management systems, documents and spreadsheets.

Claude Meetings is designed to perform more of that work in context. Anthropic says meeting results, including notes, action items and other materials, remain available in Claude after the call. Whiteboards can organize brainstorming into visual workspaces with timelines and deadlines. Sheets can create and edit collaborative spreadsheets with formulas, formatting, multiple tabs and dropdown menus. Claude Data can connect to data warehouses and customer relationship management systems to produce dashboards, written data stories and narrated videos.

Together, those products give Anthropic a potential workflow advantage. Claude is not simply answering a question or summarizing an event. It is being positioned as a layer that links discussion to business artifacts, then carries those artifacts into later decisions.

The competitive prize is the work graph

This is a direct move into territory controlled by Microsoft, Google, OpenAI and Meta. The competition is increasingly about which assistant can become the default interface for an organization’s information and processes.

Microsoft has the advantage of embedding AI into widely used productivity and collaboration products. Google can connect its assistant to Meet, Workspace and its broader cloud infrastructure. OpenAI has a large consumer and enterprise profile, while Meta can use its reach across workplace communications and social platforms. Anthropic’s opportunity is to compete on the quality of reasoning and on a more focused enterprise workflow built around Claude.

Interactive connectors are central to that strategy. Anthropic says Claude can work with applications including Adobe, Atlassian and HubSpot. Each connection can make Claude more useful, but it can also make the assistant more difficult to replace. Once an organization relies on Claude to pull information from customer systems, update workspaces and prepare operational documents, changing assistants becomes a migration project rather than a simple software switch.

The business value therefore depends on execution. Claude must reliably understand conversations, identify commitments without overstating them, use connected data appropriately and produce artifacts that employees trust. A polished demonstration can attract attention, but repeatable results across thousands of meetings will determine whether the feature becomes an enterprise habit.

Nicholas Lin, Anthropic’s product lead for knowledge work, told Axios in its report on the launch that the company had tested Claude Meetings internally for about two months and that it was participating in nearly all of Anthropic’s meetings. That internal usage supports Anthropic’s argument that the product is meant to operate as part of a routine, not as an occasional transcription utility.

The governance problem is not secondary

The same feature that gives Claude operational value also creates a more serious governance challenge. A summary can be reviewed after a meeting. An assistant that drafts a plan, updates a document or creates a timeline during the discussion may influence what participants believe was decided.

Anthropic says Claude identifies itself in the meeting and posts a notice in the chat. It also says Claude will not join calls involving external guests without administrator authorization and will leave automatically if an external guest enters. Those controls address some consent and access concerns, especially for organizations that need to separate internal discussions from conversations with customers or partners.

They do not resolve every question. Companies will need to decide what meeting data Claude can access, how long it is retained, who can retrieve the resulting materials and whether generated action items require human approval before they become commitments. Informal comments can be mistaken for decisions. A tentative deadline can become a visible project target. A statement made without the right authority can be transformed into a task that others treat as approved.

The Electronic Frontier Foundation’s senior security and privacy analyst, Thorin Klosowski, raised a related concern in an Associated Press report on AI meeting notetakers. Klosowski criticized systems that may fail to make their presence and recording clear to participants. His argument goes beyond a narrow notice requirement: people cannot meaningfully consent to an AI participant if they do not understand that it is present, what it can access or how its output will be used.

That criticism is especially relevant as assistants move from passive recording to active participation. The more Claude can do, the more important transparency becomes. A visible bot may still be misunderstood if participants assume it only takes notes while it is also querying connected systems or creating work products.

Anthropic’s next test

Claude Meetings gives Anthropic a chance to compete for a higher-value role in enterprise software. Meeting summaries are useful, but they are also becoming a common feature. The stronger commercial position belongs to the assistant that can connect meetings with the systems where work is planned, measured and completed.

Anthropic’s challenge is to make that automation dependable without allowing speed to outrun authorization. If it succeeds, Claude could become an execution layer across workplace applications. If it fails, companies may see the product as an uncontrolled source of records, tasks and commitments.

The beta is therefore testing more than a new meeting feature. It is testing whether businesses are ready to let an AI participant help decide what a conversation means, then start building the consequences before everyone has left the call.

Sources

claude.com claude.com primary
anthropic.com anthropic.com primary
techcrunch.com techcrunch.com
axios.com axios.com
apnews.com apnews.com

Originally published at spaisee.com.

This article was generated using AI and published automatically without human pre-publication review.

GPT-6 Astra edits a small model’s weights, pointing to AI debugging

Jaromir Tesar — Thu, 12 Nov 2026 14:06:01 +0000

Quick take

Qwen’s weak MATH-500 performance was attributed primarily to repetitive “thinking” loops that consumed its output limit, rather than an inability to reason mathematically.
Astra’s intervention was concentrated in two tensors, while the experiment involved tools, hypothesis testing, validation, and runs using budgets of 10, 25 and 50 forward passes.
Nathan Barry reported that a simpler linear-programming method reproduced or exceeded much of Astra’s improvement while changing roughly eight times fewer weights.
The study does not establish that direct weight editing is cheaper, more reliable, or more scalable than conventional fine-tuning.

Why it matters

For developers maintaining narrow-purpose models, the experiment shifts the proposed maintenance workflow from retraining toward diagnosing and patching internal behavior, subject to human review. What matters next is whether direct editing proves cheaper, more reliable, or more scalable than fine-tuning, and whether organizations can preserve an audit trail of changes and side effects.

When a language model gets stuck, the usual answer is more training. An MIT experiment suggests another possibility: ask a stronger AI to inspect the model’s internal settings and patch the problem directly.

VentureBeat reported that MIT researchers used OpenAI’s GPT-6 Astra to examine and modify the weights of Qwen3.5-0.8B, a small open model. After the intervention, Qwen’s score on the MATH-500 benchmark rose from 7.9% to 48.1%, while approximately 0.016% of its parameters changed.

The numbers are dramatic. The explanation is more modest, and more interesting.

The problem was not simply missing knowledge

MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) (54960889464) · ajay_suresh · via wikipedia · CC BY 4.0

In the experiment, GPT-6 Astra acted less like a chatbot answering questions and more like a technician investigating a malfunction. It read the model’s weights, tested its behavior, formed hypotheses about the failure, and then wrote changes into the network.

The researchers, Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, describe the approach in their account of weight-space programming. Their work presents large language models as systems that may be able to reason about the numerical structures inside other neural networks, rather than merely generating text around them.

Qwen’s weak MATH-500 performance was not primarily caused by an inability to produce mathematical reasoning. According to the released model documentation, the model often entered repetitive “thinking” loops. It repeated parts of its reasoning until reaching the output limit, preventing it from completing answers it might otherwise have solved.

That distinction matters. The edit did not appear to install six times more mathematical knowledge into Qwen. Instead, it helped the model stop wasting its limited response space. The improvement was therefore closer to removing a software bug than upgrading a student’s entire mathematics curriculum.

The researchers report that the change was concentrated in two tensors, or internal arrays of model values. The edited checkpoint and its benchmark results are documented in the released Qwen3.5-0.8B Astra Weight Edit model card, which also records the 0.016% parameter-change figure and the qualification about repetitive reasoning.

A promising process, not yet a practical replacement

The experiment required substantial investigation rather than a single clever instruction. Astra had to inspect behavior, use tools, test possible interventions and validate the result. The researchers also released logs and evaluation artifacts, including runs from the Qwen experiment using budgets of 10, 25 and 50 forward passes in their public experiment repository.

That makes the work feel less like magic and more like an early debugging workflow. A future AI maintenance system might diagnose why a model refuses certain inputs, loops during reasoning or performs poorly in a narrow domain, then propose a small internal repair for human review.

But the study does not establish that direct weight editing is cheaper, more reliable or more scalable than conventional fine-tuning. It also does not show that Astra was uniquely capable of finding the solution.

A critical follow-up from Nathan Barry argues that a simpler linear-programming method reportedly reproduced or exceeded much of Astra’s improvement while changing roughly eight times fewer weights. Barry’s analysis is the only opposing assessment identified for this report, and it suggests that the AI-driven search may not have been necessary.

Even so, the experiment points toward a significant shift in how models could be maintained. Instead of retraining an entire network whenever it develops a narrow weakness, developers may eventually ask one AI system to inspect and patch another.

That prospect creates practical opportunities, but also governance questions. If models can rewrite one another’s internal behavior, organizations will need records of what changed, why it changed and whether the repair introduced a new failure somewhere else. The future AI debugger may be powerful, but its most important feature could be an audit trail.

Sources

rikiyat.github.io rikiyat.github.io primary
huggingface.co huggingface.co primary
github.com github.com primary
newruntime.com newruntime.com primary
venturebeat.com venturebeat.com

Originally published at spaisee.com.

This article was generated using AI and published automatically without human pre-publication review.

Gemini Robotics 2.5 pushes robots toward factories, but reliability still lags

Jaromir Tesar — Thu, 12 Nov 2026 13:09:03 +0000

Quick take

DeepMind evaluated Gemini Robotics 2.5 on more than 200 tasks, including pipetting, folding garments and tying bows.
At Mercedes-Benz, an Apptronik Apollo 2 humanoid achieved a 96% success rate handling automotive parts in a kitting workflow.
Ryan Whitwam called the Mercedes result short of human-level reliability, while Carolina Parada said generalization may hinder 99.9% reliability.
DeepMind introduced ASIMOV-Jailbreak, an open benchmark testing whether robots reject unsafe instructions and resist adversarial prompts.

Why it matters

For manufacturers, the shift is from narrowly programmed demonstrations toward more adaptable robots that can reduce the need for separate control systems, though buyers may still restrict them to low-risk work or keep humans involved. The next test is whether Google and its partners can cut supervision, downtime and integration costs while delivering predictable performance and real-world safety.

Google DeepMind’s latest robotics model improves dexterity, speed and task success, but its 96% factory result highlights the commercial gap between impressive demonstrations and the consistency manufacturers require.

A meaningful step into physical work

Google’s reported gains versus its predecessor and Mercedes-Benz kitting result

Google DeepMind said in its announcement that Gemini Robotics 2.5 was evaluated on more than 200 tasks, including activities such as pipetting, folding garments and tying bows. Compared with its predecessor, the company reported success-rate improvements ranging from 30% to 100%, alongside speed gains of 20% to 30%.

Those figures matter because robotics companies have historically struggled to build systems that can generalize beyond tightly controlled demonstrations. A robot that can perform one assembly task repeatedly may be useful, but a commercially valuable system must handle variation in objects, positioning, lighting and instructions without extensive reprogramming.

Google’s strongest industrial result came from a proof of concept involving an Apptronik Apollo 2 humanoid robot at Mercedes-Benz. DeepMind said the system achieved a 96% success rate while handling automotive parts in a kitting workflow. For the AI industry, that is evidence that large multimodal models are becoming more capable of controlling physical machines in environments with economic value.

For manufacturers, however, the number has a different meaning. A 96% success rate can still produce too many failures when a robot performs thousands of operations, particularly if a mistake damages a component, stops a production line or creates a safety risk.

The final percentage points are the market

Ars Technica reported that Ryan Whitwam viewed the Mercedes result as short of human-level reliability. The report also quoted DeepMind’s Carolina Parada, who said that generalization may come at the expense of reaching 99.9% reliability in every application.

That tradeoff is central to the commercial strategy behind general-purpose robotics. Broader capability can reduce engineering and deployment costs because customers may not need to build a separate control system for every task. But if that flexibility introduces even occasional failures, buyers may limit the robot to low-risk workflows or retain human oversight.

This creates a difficult business calculation for Google and its robotics partners. A general model can offer a wider addressable market than a specialized industrial system, but the value of that market depends on whether customers can trust the system enough to use it at scale. The competitive advantage will not come from demonstrations alone. It will come from reducing supervision, downtime and integration costs while maintaining predictable performance.

Latency adds another constraint. Ars Technica noted that Gemini Robotics 2.5 still runs on cloud infrastructure, which can make response time a concern. In factory operations, connectivity and inference delays can undermine the benefits of faster physical execution. Cloud dependence may also raise questions about data security, operational resilience and the cost of running models continuously.

Safety benchmarks are useful, not decisive

DeepMind also introduced ASIMOV-Jailbreak, an open benchmark designed to test whether robotic systems reject unsafe instructions and resist adversarial prompts. The Asimov Agentic Safety Evaluation dataset includes the dataset files and evaluation scripts for the physical-safety and jailbreak testing suite.

The benchmark gives developers a shared way to measure whether a model follows instructions that conflict with physical constraints or safety policies. That is strategically important as robotics systems become more autonomous and are exposed to natural-language commands from workers, software agents and potentially untrusted users.

Still, New Runtime said its coverage that benchmark scores alone cannot establish that a robot is safe in deployment. The publication argued that testing must be combined with runtime controls, hardware safeguards and validation in real-world conditions.

The supplied reception of Gemini Robotics 2.5 is critical, and no opposing view was found in the provided coverage. That criticism does not erase the progress. It identifies the business test Google must now pass: turning better generalization into factory-grade consistency.

Gemini Robotics 2.5 may give Google a stronger position in the race to supply the intelligence layer for commercial robots. But until reliability approaches the standards of industrial automation, the model is more likely to expand pilots than replace established production systems.

Sources

deepmind.google deepmind.google primary
huggingface.co huggingface.co primary
newruntime.com newruntime.com primary
arstechnica.com arstechnica.com

Originally published at spaisee.com.

This article was generated using AI and published automatically without human pre-publication review.