DEV Community: Chandan Kumar

Gemini in CarPlay ships as a navigation feature, not an assistant

Chandan Kumar — Wed, 21 Oct 2026 18:55:09 +0000

Google's conversational Gemini is now live in Apple CarPlay for US drivers, and the detail that matters for anyone building in-car voice is where it landed: inside navigation, not as a standalone assistant.

What actually shipped

According to The Verge, Google brought Gemini to Android Auto first, and is now extending the conversational navigation experience to CarPlay users in the US, with a global rollout also underway. Google hasn't published specific regional availability yet. The gap between the Android Auto launch and the CarPlay launch is roughly twelve months.

That twelve-month lag is the headline for platform planning. If your roadmap assumes a voice feature you ship on Android Auto is a quarter away from parity on CarPlay, this is a concrete counterexample from the company with the most leverage over both platforms.

Why the navigation frame is the interesting part

CarPlay's app model doesn't offer a general-purpose assistant bucket. The categories Apple exposes are deliberately narrow, and maps is one of the few places a third party can hold a persistent, voice-forward surface in the car. So an assistant that wants to be present while someone drives has to arrive through the navigation template or not at all.

That constrains the product in useful ways. A conversational layer wrapped around routing has an obvious job: interpret an ambiguous destination, chain a waypoint, answer "what's near the next exit" without the driver tapping anything. Every one of those is scoped to a session that already exists. You aren't designing an open-ended chat client for a windshield; you're designing utterances that resolve into route mutations.

It also means the interesting failure modes are different from a phone assistant. Partial addresses, noisy cabins, and users who say "the one by the Target" are the test suite. Latency matters more than breadth, because a wrong turn costs more than a wrong answer.

What to take from it

Ship the assistant inside a surface the platform already blesses, and let the host app's job define the intent set.
Treat cross-platform voice parity as a multi-quarter problem, not a port.
Budget for ambiguity resolution against real map data, not generic chat quality.

The competitive question isn't whether Gemini sounds better than Siri in a car. It's whether Google can keep shipping conversational routing on a platform where it doesn't control the app categories, and how much of that twelve-month gap is structural versus scheduling.

Source: The Verge

Source: https://www.theverge.com/transportation/1015281/gemini-ai-apple-carplay-voice-google-maps

Google has joined Amazon in attacking Nvidia's stronghold with AI app developers

Chandan Kumar — Tue, 20 Oct 2026 19:47:02 +0000

The most useful detail in Google's new developer tool for moving PyTorch workloads onto its own silicon is that Google is now competing on porting cost instead of benchmark charts. As Julie Bort reports for TechCrunch, Google has joined Amazon in releasing a tool aimed at making it easy for PyTorch users to move their apps off Nvidia GPUs.

That framing matters more than any spec sheet, because it attacks the part of Nvidia's position that actually holds customers in place.

The moat was never really the chip

PyTorch is where model code gets written, and CUDA has been the default compilation target for so long that "runs on PyTorch" and "runs on Nvidia" quietly became the same sentence in most teams' heads. The cost of leaving isn't the price of a new accelerator card; it's the quarter of engineering time spent rewriting custom kernels, re-checking numerics, and re-tuning wherever the profiler now points somewhere unexpected.

A vendor that ships tooling for PyTorch users specifically is saying: your code, as written, is the input. That's a different sales motion than "here are our TOPS." Amazon got there first with its own developer path off Nvidia hardware. Google following means the pitch is no longer a one-off experiment a platform team has to defend alone in a roadmap review.

What to test before you commit a cluster

If a porting path is on the table, the useful diligence is unglamorous and mostly about where the abstraction leaks:

Coverage for your custom ops and any fused kernels you hand-wrote, not just the ops in a tutorial model.
Numerics parity on your actual loss curves, at your actual batch sizes, not a cosine-similarity check on one layer.
Multi-node behavior, since interconnect behavior is usually where a smooth single-node port falls apart.
The escape hatch: what happens when an op isn't supported and you need to fall back, and how much of the graph that fallback drags with it.

None of these questions get answered by a migration guide. They get answered by porting one real training job and one real inference service and timing the diff.

Why the second vendor matters

The practical consequence of Google joining Amazon here isn't that Nvidia loses a benchmark. It's that "we want a second source for accelerators" becomes a defensible procurement position rather than a science project. For anyone who has argued that single-vendor dependency on compute is a business risk, the argument just got cheaper to make — and the honest answer to "how long would a port take?" is now a number you can measure instead of guess.

Source: Google has joined Amazon in attacking Nvidia's stronghold with AI app developers

Source: https://techcrunch.com/2026/10/20/google-has-joined-amazon-in-attacking-nvidias-stronghold-with-ai-app-developers/

Why Tolan's $30M bet on planning over chat changes how we build AI companions

Chandan Kumar — Mon, 19 Oct 2026 19:21:17 +0000

Most AI companions are chat-first: they ask how your day went, offer sympathy, and maybe generate a poem. Tolan raised another $30 million, per TechCrunch, because it took a different route — planning-first. With over 250,000 active users, Tolan's traction suggests that the killer use case for an AI companion isn't conversation; it's helping someone navigate the concrete steps of their life.

What Tolan does differently

Tolan is an AI companion app designed to help users navigate life. The key insight is that the app focuses on actionable planning rather than open-ended dialogue. Instead of asking "How are you feeling?" it might help you break down a career move, plan a move to a new city, or map out the steps to start a side project. The companion acts as a structured guide, not a therapist or a friend.

This matters because the market is crowded with chat-based companions that struggle to retain users after the novelty wears off. Tolan's approach ties the AI to outcomes — completing a task, making a decision, or executing a plan. That creates a repeatable value loop: the user returns because the AI helped them do something, not just because it listened.

What the funding means for the AI companion landscape

The $30 million round signals that investors see planning as a defensible niche. Most LLMs are good at generating text; few are good at generating a realistic, step-by-step plan that accounts for a user's specific constraints. Tolan likely had to build custom scaffolding — retrieval-augmented generation for personal context, task decomposition, and follow-through tracking — on top of a base model to make this work.

For developers building similar products, the takeaway is clear: don't race to build the best chatbot. Race to build the AI that helps someone finish something. The companion market may ultimately be won by the team that makes the user's life more navigable, not more conversational.

Source: TechCrunch

Source: https://techcrunch.com/2026/10/19/tolan-snags-30m-for-its-ai-companion-that-helps-users-navigate-life/

Disney's Character.AI Bid Is a Hiring Signal, Not a Done Deal

Chandan Kumar — Mon, 19 Oct 2026 05:58:56 +0000

The detail that changes how to read Disney’s Character.AI move is the deal structure: the workforce transfer is substantial, but it is a stalking horse bid, so another company could still come forward with a higher offer. According to TechCrunch’s Anthony Ha, a filing says Disney has agreed to hire the majority of Character.AI’s workforce, and the “vast majority” of employees who were offered jobs accepted them.

What the filing confirms

The concrete signal is people, not a model launch or a benchmark result. Disney is absorbing most of the Character.AI team through a hiring arrangement tied to a bid. The filing also states that most of the employees who received offers said yes. That is a stronger commitment signal than a vague talent announcement, but it is not the same as a closed acquisition.

The distinction matters because AI companies are often evaluated through their models, compute, and product surface. Here, the transferable asset being described is the workforce. For anyone tracking the consumer AI companion space, the useful question is not just “Who owns Character.AI?” but “Which team, product, and contract obligations move if this bid survives?”

Why a stalking horse bid matters to builders

A stalking horse bid sets an initial offer that others can beat. That means the Disney deal may be the floor, not the final price. For developers and technical buyers, this should change the risk calculation. If your roadmap depends on Character.AI-powered features, an API integration, user data export, or moderation workflows, do not treat the Disney hiring news as a vendor-stability guarantee.

Separate the people signal from the asset signal. Hiring intent can survive a topping bid; product continuity and contract terms may not. The filing makes the workforce outcome look likely, but the sale process leaves room for a competing offer.

What to do with the signal

Put a review date on any Character.AI dependency and map whether it is people-dependent or contract-dependent. If a renewal is coming up, ask for language that survives an ownership change. If you are evaluating the space, note that Disney’s move is a talent-heavy bet with an open auction structure, not a completed integration.

The headline is not “Disney buys Character.AI.” It is “Disney is hiring most of the team through a bid that can be topped.” Build your dependency map for both outcomes.

Source: TechCrunch

Meta removes scraping language from commercial terms after backlash

Chandan Kumar — Sat, 17 Oct 2026 19:53:34 +0000

Meta just quietly removed language from its commercial terms that would have given it a broad license to scrape content — including from news sites — to train its AI models.

If you build anything that relies on Meta's AI infrastructure or that publishes content on the open web, this change matters. Here's what happened and why you should care.

What they removed

Meta's commercial terms previously included this phrase: "Meta may use any content from such domains, web pages, or URLs to provide, develop, improve, and personalize our products and services."

That language is now gone, according to a report from The Verge. The removal follows public backlash from publishers and site owners who interpreted the clause as giving Meta permission to scrape their content for AI training without consent.

Why it matters for developers

If you publish content, this change removes one explicit legal hook Meta could have used to scrape your site for AI training without a separate agreement.
If you build on Meta's platforms, the terms around how your users' content gets used for model training just got more restrictive. You can no longer assume Meta has a blanket license to consume everything on commercial domains.
If you're evaluating AI vendors, this episode is a reminder to read the fine print on data usage clauses. A single sentence in commercial terms can override privacy promises elsewhere.

The bigger picture

Meta's move is a response to growing pushback from publishers who watched their content get scraped by AI companies with little recourse. But it's also a signal: the terms-of-service battle over training data is far from settled. Other platforms may follow or backtrack depending on regulatory pressure.

For now, if you run a site that Meta might want to scrape, this change gives you slightly more clarity. But the fundamental tension between open web content and AI training remains unresolved. Keep watching the terms pages.

Source: The Verge - Meta removes language that may have allowed it to scrape news sites

Source: https://www.theverge.com/tech/1013409/meta-removes-language-that-may-have-allowed-it-to-scrape-news-sites

Robotics is stuck in its Lego era — here's what's missing

Chandan Kumar — Fri, 16 Oct 2026 21:08:41 +0000

Robotics hasn't had its ChatGPT moment yet, and according to Amulya Vishwanath — who spent seven years leading product marketing for Nvidia's robotics platforms — the reason is that the entire field is still rebuilding its basic building blocks.

Vishwanath describes robotics as being in its "Lego era." The actuators, robotic hands, software stacks, and safety standards that would let you snap together a general-purpose robot simply don't exist yet. This directly contradicts the hype that a single foundation model is about to unlock human-level dexterity and reasoning in any robot.

Where the bottleneck really is

Most developers assume the hard part of robotics is the AI. If you can just train a big enough model on enough video, the robot will figure out how to fold laundry or open a door. But Vishwanath's view from inside Nvidia suggests the harder problems are mechanical and systemic. The physical components — motors, grippers, sensors — are still being designed and redesigned with no standard form factor. Software frameworks for control and simulation are fragmented. Safety certification processes haven't been updated for the age of learned behaviors.

This matters for anyone building on top of robotics platforms today. You're not just integrating an API; you're assembling a system where every layer is still in flux. The foundation models we have (like Google's RT-2 or Nvidia's own Isaac tools) are impressive, but they can't compensate for a robot arm that doesn't have the precision to match its neural net's predictions.

What a ChatGPT moment would actually require

A true inflection point in robotics would mean:

Standardized, off-the-shelf hardware that works out of the box with any software framework
A general-purpose foundation model that can be dropped into any robot body without retraining
Clear regulatory guardrails so that deploying a robot in a hospital or warehouse doesn't require a custom legal review

None of those exist yet. The industry is still building the Lego bricks. Vishwanath isn't saying it won't happen — but the timeline is longer than the current hype cycle suggests.

Source: TechCrunch (2026-10-16) — An Nvidia insider-turned-VC on why robotics hasn't hit its ChatGPT moment, yet by Theresa Loconsolo

Source: https://techcrunch.com/video/a-nvidia-insider-turned-vc-on-why-robotics-hasnt-hit-its-chatgpt-moment-yet/

ChatGPT’s Intelligent UI turns responses into interactive apps, not just text

Chandan Kumar — Wed, 07 Oct 2026 19:18:36 +0000

OpenAI’s new Intelligent UI feature for ChatGPT fundamentally changes what a chatbot response can be: it now ships diagrams, charts, tappable buttons, and fillable forms alongside its text, turning a single reply into an interactive mini-app.

This isn’t just a richer markdown renderer. The update, rolling out alongside GPT-6, lets ChatGPT compose responses that combine natural language with live, actionable widgets. Instead of describing a bar chart in prose, the model can embed the chart directly. Instead of listing steps, it can present a form with checkboxes, dropdowns, and submit buttons.

How it changes the interaction model

Traditional chatbot replies are linear text blocks. Intelligent UI breaks that by letting the model output a structured layout that mixes static content with interactive elements. The key shift is that the user doesn’t have to leave the chat to interact with data visualizations or simple workflows.

For example, if you ask ChatGPT to compare quarterly revenue trends, the response might include an embedded line chart with hover detail, a table of raw numbers, and a button to export the data. If you ask for a project timeline, you might get a Gantt chart and a form to adjust dates.

What it means for developers

This update reduces the need to build custom front-ends for common data exploration tasks. If your team uses ChatGPT as a knowledge interface, Intelligent UI can now present complex information in a consumable, interactive format without requiring a separate dashboard or API call.

The feature also suggests a new pattern for conversational UIs: treat the chat window as a canvas for lightweight, composable widgets rather than a scrolling log of text. OpenAI hasn’t published the widget SDK or schema yet, but the model clearly has fine-grained control over layout and interaction type.

One caveat

The feature works across all ChatGPT tiers, but its effectiveness depends on the model’s ability to choose the right visual or interactive element for the query. Early examples from OpenAI show charts and forms, but edge cases—like overly complex dashboards or conflicting interactive elements—are likely. Expect iteration on the heuristics that decide when to use a widget versus plain text.

Source: The Verge — ChatGPT’s ‘Intelligent UI’ update

Source: https://www.theverge.com/ai-artificial-intelligence/1007276/openai-chatgpt-intelligent-ui-gpt-6

Hark's AI assistant runs locally as an operating system layer, not an app

Chandan Kumar — Tue, 06 Oct 2026 19:14:21 +0000

Hark's new personal assistant is an operating system from the future, not just another chatbot app.

That distinction changes how you should think about building AI features. Most assistants run as a web service or a mobile app layer, sending prompts to a cloud model. Hark flips the architecture: the assistant operates at the OS level, processing data locally and only reaching out to the cloud when the user explicitly authorizes it.

What makes Hark different

Hark is an AI lab that just launched a personal assistant designed to compete with Muse, Dots, and Instinct. The key design choice is privacy-by-default: the assistant runs on-device, intercepts system events, and learns from your files, calendar, and messages without uploading them.

From a developer perspective, this means:

The assistant has access to the full OS event stream — file changes, notifications, app launches — not just a sandboxed API.
All inference and storage happen on-device unless the user opts into a cloud query.
The model can be updated without a full OS patch, similar to how browser engines get point releases.

How this affects your integration strategy

If you're building tools that need to coexist with Hark's assistant, you have two paths:

Local-first APIs: Your app should expose structured data through system share sheets, file providers, or notification categories. Hark's assistant consumes these OS-level channels, not a custom SDK.
Privacy-preserving hooks: Because the assistant does not send data out by default, you cannot rely on a cloud callback to enrich user actions. Design your app to do heavy lifting locally or precompute summaries.

Hark's approach is a bet that users will trade cloud-powered features for privacy guarantees. For now, the assistant is limited to macOS and iOS, with an Android version in development.

If you are evaluating which AI assistant ecosystem to build against, Hark's local-first architecture is the most developer-friendly for privacy-conscious apps — but you lose the convenience of cloud-based model chaining.

Source: TechCrunch

Instinct puts its AI agent in group chats, and non-users can join the thread

Chandan Kumar — Mon, 05 Oct 2026 18:56:42 +0000

The hard part of multi-user AI agents isn't the model, it's deciding which slice of a person's private context a shared agent is allowed to see. Instinct's new group chats are a live test of that.

What actually shipped

Instinct is launching group chats where friends use its AI agent together on tasks like planning trips, organizing carpools, and coordinating events. The detail that matters for anyone building in this space: friends without an account can take part. The agent, not a signup form, becomes the interface to the product.

That's a distribution move, and a good one. If a non-user can sit in the thread and still get value from the agent, onboarding happens inside shared context instead of behind a wall. But it also means the agent is now serving people with different context, different trust levels, and different stakes.

Two agents wearing one name

Instinct says personal accounts remain separate, with permission required before personal agents share information or take action. Read that literally and you have two distinct components, not one assistant with a group mode:

A personal agent holding private context — preferences, history, whatever it knows about you
A group-scoped agent that only receives what has been explicitly released into the thread

The boundary between them is the product. Everything else — the itinerary, the carpool roster — is downstream of that boundary holding.

Model consent at the thread, not the user

If you model permission as per-user, per-action, a five-person thread turns into a permission-prompt generator. Nobody will approve eleven dialogs to schedule a Saturday. Thread-scoped grants are more workable: this thread gets my availability windows, not my location; this thread can write to the shared plan, not to my calendar.

Revocation has to be visible to the whole thread, not buried in settings. In a group, one person's leaked detail poisons trust for everyone, and a silent revoke reads as either a bug or a betrayal. Make the release of information an event the thread can see.

Why this class of task is the hard case

Planning trips, carpools, and events are coordination problems where the bottleneck is agreement, not generation. An agent that drafts a perfect itinerary the group won't accept has failed. The permission layer is what lets the agent act on someone's behalf without asking them to narrate their whole life to their friends first.

That's the design lesson worth carrying into other multi-user agent products: build the consent gate as a first-class object in the data model, with a clear scope and a visible revocation path. The shared agent's answers are only as good as the context it's allowed to have.

Source: TechCrunch, "Instinct brings its AI agent to group chats, even for friends without an account," Sarah Perez, 2026-10-05.

Source: https://techcrunch.com/2026/10/05/instinct-brings-its-ai-agent-to-group-chats-even-for-friends-without-an-account/

AI Deepfake Defense: NJ Lt Gov Uses Synthetic Video to Deny Harassment

Chandan Kumar — Sun, 04 Oct 2026 18:51:49 +0000

AI-generated media is no longer just for memes or movie cameos — it's now being used as a legal defense by a sitting politician.

Former New Jersey Lieutenant Governor Dale Caldwell, who resigned September 25 after an investigation found he sexually harassed a staffer and repeatedly violated ethics rules, has taken an unusual tactic to clear his name: he's using synthetic video footage that appears to show him making different statements during the alleged incidents.

How the AI Defense Worked

During a media interview on NJ Spotlight News, Caldwell presented what he claimed was AI-generated video evidence that contradicted the investigation's findings. According to reporting by The Verge, the former Lt. Governor argued that the original recordings had been manipulated and that his AI-generated recreation showed the "real" version of events.

The technique is a novel application of deepfake technology — but not in the way most developers think about it. Instead of creating fake content to harm someone, Caldwell is using synthetic media to argue that real evidence against him is fake. This inversion of the deepfake threat model is something security and AI practitioners should watch closely.

Technical Implications for Developers

This case raises several practical concerns for anyone building or deploying AI media generation tools:

Authentication becomes adversarial: If bad actors can plausibly claim real footage is AI-generated, the burden shifts to proving authenticity. This is a fundamentally harder problem than detecting fakes.
Latent space manipulation: The AI models used to generate these defenses likely rely on diffusion or GAN-based architectures. Developers should consider how their model's latent space could be exploited to produce "counter-evidence" that mimics real footage.
Metadata stripping: The synthetic videos may lack proper provenance metadata (C2PA, etc.), making it harder for forensic tools to distinguish them from authentic recordings.

What This Means Going Forward

This isn't a hypothetical scenario from a conference talk — it's happening now in a high-profile political case. The investigation that led to Caldwell's resignation was conducted by the state's ethics board and found "credible evidence" of harassment. Whether the AI defense holds up in court remains to be seen, but the precedent is set: synthetic media is now a tool for disputing real evidence, not just creating fake content.

For developers working on deepfake detection, verification, or AI-generated content pipelines, this case underscores the need to build systems that can prove authenticity as robustly as they can detect fakes. The arms race just got more complicated.

Source: The Verge

Source: https://www.theverge.com/ai-artificial-intelligence/1004549/well-if-ai-said-it-it-must-be-true

AWS drops NDAs for data center energy use, but the gap remains

Chandan Kumar — Sat, 03 Oct 2026 19:10:38 +0000

AWS no longer requires NDAs for data center electricity contracts. That is the headline from AWS CEO Matt Garman’s recent response to community backlash, as reported by TechCrunch. The change signals a shift, but for developers building on AWS, the real question is whether this changes anything about how you evaluate a cloud provider's environmental footprint.

What actually changed

Garman stated that AWS has stopped using non-disclosure agreements in contracts with local utilities and data center operators. Previously, these NDAs prevented communities and local governments from knowing how much power a specific AWS data center consumed. The move follows years of criticism that cloud providers operate in an energy transparency vacuum.

Why this is still not enough

Dropping NDAs from utility contracts does not mean AWS is publishing per-datacenter energy data. It means the utility is no longer contractually gagged. The data may still be hard to find, unstandardized, or buried in public utility filings that vary by jurisdiction. For a developer who wants to know the carbon intensity of their workload running in us-east-1, the practical path to that number remains unclear.

What practitioners should watch for

If you are evaluating cloud providers on sustainability, look for the following concrete signals:

Per-region or per-availability-zone energy usage reports published on a regular cadence
Third-party audited carbon accounting (not just self-reported offsets)
Real-time or near-real-time grid mix data for the regions you use

AWS’s own customer carbon footprint tool gives you a monthly estimate, but it is an average. For latency-sensitive or batch workloads that can shift to greener hours, averages are not actionable.

The bottom line

This is a step toward transparency, not a finished solution. AWS removed a contractual barrier, but the data itself is still not surfaced in a developer-ready way. Until per-zone energy data is available via API or dashboard, the trust gap remains.

Source: https://techcrunch.com/2026/10/03/amazon-responds-to-data-center-backlash-says-it-no-longer-uses-ndas

Source: https://techcrunch.com/2026/10/03/amazon-responds-to-data-center-backlash-says-it-no-longer-uses-ndas/

Meta open sources the code to build your own Muse AI companion gadgets

Chandan Kumar — Fri, 02 Oct 2026 22:04:24 +0000

Meta just open-sourced the core code that powers its Muse AI agent, and now you can build your own physical gadgets around it.

What you get

Meta released the software stack that runs Muse, its conversational AI agent. The code is available so anyone can load it onto custom hardware. Meta suggests several reference builds: putting Muse on a color E Ink display for showing reminders, or embedding it into an HDMI stick to display on a large screen.

This is different from just calling an API. You get the actual agent runtime, meaning your device can run Muse locally or with minimal cloud dependency, depending on your build.

How to use it

Download the open source code from Meta's repository.
Pick a compatible display (color E Ink is confirmed) or an HDMI output board.
Flash the firmware and configure your agent preferences.
Connect to power and a network — Muse handles the rest.

The HDMI stick approach is particularly interesting for developers. You can plug it into any monitor or TV and have a persistent AI assistant visible at all times, without needing a dedicated screen.

Why this matters for builders

This is a shift from the usual "we ship a device, you buy it" model. By open-sourcing the code, Meta lets you control the hardware, the form factor, and the privacy boundaries. You can build a Muse gadget that only connects to your local network, or one that surfaces specific data from your own services.

For now, the builds focus on display-only interaction — showing reminders and information. But the underlying agent can handle voice and text conversations, so future hardware could include microphones and speakers.

If you've been waiting for an AI companion you can actually own and modify, this is the first major release that makes that possible without reverse-engineering a closed device.

Source: The Verge — Meta open sources code to let you make Muse AI gadgets

Source: https://www.theverge.com/tech/1004330/meta-muse-ai-gadgets-home-link