The Most Useful New AI Features and Tools to Try

August 28, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
August 28, 2026

HEADLINES

NVIDIA Buys Hugging Face

Confirmation landed late Wednesday night, right after the earnings call: NVIDIA has agreed to buy Hugging Face for $12.9 billion. On $150 million in annualized revenue that works out to roughly 80x revenue, against something like 15x for SpaceX's $60 billion acquisition of Cursor — but NVIDIA is very much not buying Hugging Face for the revenue. Stacked next to the $6 billion Poolside licensing deal that brought in around a hundred veteran researchers, the Perplexity partnership, and NeMo-Tron's climb from tech demo to real contender, this looks like NVIDIA acquiring its way into a full-stack open business model. It's also defensive, and not just against custom silicon: when Meta or Mistral or DeepSeek release weights, millions of developers fine-tune and serve them overwhelmingly on NVIDIA GPUs through CUDA, which is exactly the world where a diversity of lighter, cheaper, more customizable models runs the day-to-day. The usual neutrality worries are out there — Hugging Face's value is partly that it isn't seen as pushing anyone toward one vendor — but the optimism around this one is unusually high, which says something about the trust NVIDIA has built in the AI community.

The Information Nvidia Agrees to Buy Open Source AI Platform Hugging Face For $12.9 Billion
Aakash Gupta (X) The defensive read: every biggest customer is building an escape route, and open models are the counterweight
Gianluca (X) NVIDIA owns the GPUs, CUDA owns the developers, Hugging Face owns the open source ecosystem
Foombler (X) They don't want to just sell the picks and shovels, they want to own more of the AI economy built on top of them
Todd Saunders (X) Hugging Face is valuable partly because it's perceived as neutral, and an acquisition could threaten that
Pratham (X) What happens to open source AI when the biggest players start buying the places where it lives
Eric Hartford (X) Happy for the exit, but HF is no longer hardware agnostic — an opening for a new standard bearer
Conor Dart (X) Can't think of a better company to buy Hugging Face, and it aligns with what NVIDIA is doing as a whole
Ramez Naam (X) Such a great push towards plural AI — it reduces the odds any one lab or nation monopolizes the frontier

NVIDIA Blows the Doors Off

Revenue growth accelerated to 106% year over year, 21 points higher than the growth rate at the end of Q1, with quarterly sales of $96.2 billion putting NVIDIA tantalizingly close to the $100 billion quarter that only five other companies have ever hit. Guidance still points to a slight slowdown — 89.5% for Q3 and 70% for fiscal 2027 on supply chain snarls — and the real red flag was the balance sheet, with free cash flow halving to $21 billion, accounts receivable up 50%, and days sales outstanding stretching from 45 to 60 as major customers ask for more time to pay on "extended payment terms on large multi-quarter agreements with certain investment-grade customers." NVIDIA also confirmed it is officially back in China with a small volume of H200 sales under the new export rules, well short of what the administration permits. Hard to overlook the raw numbers, though: sales have nearly doubled since last summer, and Amazon just ordered another two million chips against an annual total believed to be under ten million.

NVIDIA Pauses Neocloud Revenue Sharing

One interesting sub-story given the unique role NVIDIA now plays across the ecosystem: the company has paused the revenue-sharing deals it announced less than two months ago, where neocloud customers could receive credit support in exchange for a cut of the profits. Sources say the program was put on hold after employees raised concerns about antitrust scrutiny and about how much control NVIDIA would be exerting over customer businesses — reportedly including approved-customer lists for the chips and a stated preference for diverse rather than single-tenant customer bases. The precise reason couldn't be learned and the program could return with tweaks; NVIDIA denies the reporting outright, saying the July model "is still in place and continues to evolve due to high demand." Worth noting the scale already committed: $36 billion in deals, typically running six years, in under two months.

Salesforce Leads the SaaS Comeback

Six months ago the narrative was that software was done for and companies would vibe code their own SaaS replacements. That level of disruption did not play out, and what's actually happening is that AI adoption inside traditional software products is driving a revenue boost — Agentforce is now on track for $1.5 billion this year, up from a $1.2 billion forecast in Q1, on overall quarterly sales of $11.5 billion growing at 11%. AI revenue is still a modest slice of the picture, but the stock jumped 22% on Thursday, leaving Salesforce about 5% away from filling the entire SaaSpocalypse drawdown, and ServiceNow, Figma, Atlassian, CrowdStrike and Okta have all staged comebacks without needing blockbuster numbers. As KeyBanc's Jackson Ader put it, "we're all coming around to the idea that things are going to be more durable and are not going away. That's why you see Salesforce come out with fine but not stellar results, but you see a really outsized reaction." Benioff took the victory lap he'd earned, ticking through every skeptic prediction — declining seats, customer attrition, eroding pricing power — that didn't happen, and landing on a six-fold surge in agentic use of the platform via MCP calls: "Apps are not dying."

The App Layer's Revenue Boom

Bringing the booming revenue story down to the app layer: Cognition has hit $900 million in annualized revenue, more than a 3x increase since the start of the year for the creators of Devin. It's a similar story across the board, with Higgsfield tripling to $700 million, Perplexity tripling to $750 million, OpenEvidence doubling to $300 million, and even Manus claiming a $400 million run rate after the messy unwind of the Meta acquisition. And while the totals remain far behind the US, leading Chinese AI companies are seeing the same shape of growth — DeepSeek has generated $70 million in revenue this year, a 10x increase on its full-year 2025 numbers.

No FINRA for AI

The executive order that would have created a self-regulatory body for the AI industry has stalled out after drafts were quietly circulated in recent weeks. The idea gathered steam after DeepMind chairman Demis Hassabis formalized it in a July blog post — a body modeled on FINRA, requiring labs to share notes on security issues, conduct safety testing and write the rules that govern the industry — and picked up public support from Treasury Secretary Scott Bessent. Why it stalled isn't clear, though opposition from former White House AI czar David Sacks is a candidate: on last weekend's All-In, he called FINRA for AI a "horrible idea" and a Trojan horse for an open source model ban, on the theory that frontier labs would push for standards applying equally to all models that open developers couldn't possibly meet. He also pointed out that FINRA doesn't have a good reputation for promoting competition and innovation, with startups entering financial services viewing it as largely concerned with using regulation to protect incumbents. Whatever the case, for now: no FINRA for AI.

100+ Companies Call for Collective Cyber Defense

Self-regulation isn't dead, it just moved: OpenAI, Anthropic and more than a hundred other companies across tech, finance and industry — Google, Microsoft, AWS, Visa, General Motors, PwC — have signed an open letter calling for an urgent new approach to cyber defense. The warning is that there's a limited window, that AI-enabled attacks will become far more widespread and sophisticated in the coming months, and that hospitals, water treatment plants and the infrastructure powering the internet are all exposed. The proposed response is collective: better AI tools in the hands of defenders, coordination across industries, resourcing for critical infrastructure security teams that are chronically under-resourced, and leadership treating this with the urgency of an active incident. In the wake of the Hugging Face incident this reads as the industry's answer, and Sam Altman positioned security above competition — "we are happy if you want to work with us or any of our competitors or partners, but please take this moment seriously." Quietly, one of the bigger points here may be pressure on the administration to stop limiting access to strong cyber models.

MAIN STORY

The Most Powerful New AI Features and Tools to Try

Just as it's nearly impossible to keep up with the onslaught of AI news, it's extremely difficult for any one person to track every new feature, model and tool that ships in a given week. Attention naturally goes to the big model announcements that might fundamentally shift how we interact with AI, but it's worth periodically taking inventory of the feature updates, quality-of-life upgrades and new capabilities to see which ones are actually relevant to how you work. What follows is roughly a dozen releases from the last seven days that are useful and practical right now.

BRAND NEW STUFF

Claude Cowork
"AI that opens the tab and gets the job done"
Cowork now has a built-in browser, so if a task involves filling out a form or interacting with a web app, a browser window opens alongside the session for Claude to use. Nothing to install, built into the desktop app, and separate from your own browser and logins — it's its own browser, not a takeover of yours. For many this was a big old finally, and while Anthropic is late here relative to ChatGPT Work and the obvious inspiration in Grokbot, combined with everything else shipping this month it adds up to a real inflection: the line between AI assistant and AI operator just got a lot thinner.

ChatGPT Work
"Its own computer in the cloud"
ChatGPT Work can now use its computer and browser to sign in to websites on web and mobile without ever seeing your username or password. The pitch list runs about two dozen items long — set up utilities for a new apartment, book a DMV appointment, check reimbursement costs through your insurance, prepare car registration renewal paperwork, save apartment listings, schedule a package pickup. Given the ongoing conversation about consumer AI versus B2B, it's interesting that these very life-oriented tasks are being pitched as the new capability for the work product.

Temporary Chats That Stick
Temporality and permanence, switchable after the fact
A smaller quality-of-life upgrade for anyone with privacy concerns who defaults to temporary chats: you can now personalize a temporary chat with your existing memories, custom instructions and plugins, and save it to your history if it turns into something worth keeping. Temporary chats still don't create new memories and stay out of the sidebar unless you choose to save them.

Multiple Gmail Accounts
Punching way above its weight class in newsworthiness
Multiple Gmail accounts are now available in ChatGPT plugins — instead of connecting one Gmail, you can connect several. This has long sat on power users' wish lists, given that the normal course of business is juggling half a dozen emails that all matter, and it was one of the most-praised features of the new Grok, which doesn't constrain connectors to a single Gmail or Slack. This is one where it doesn't matter who did it first as long as they all do it.

Hermes
Keeping pace with every major harness, within days
For anyone who got hooked on the flexibility of something like OpenClaw, the Hermes agent is worth watching: the team is clearly tracking everything the harness and model companies ship, and anything useful finds its way over extremely quickly. Thursday brought real-profile browsing, where the agent acts with your logins from a managed copy of your existing Chrome profile — following bot mode, introduced days after the Grokbot announcement, which turns agent profiles into named bots with their own role, model, memory, skills and picture. The pitch is control: its own remote computer at a few dollars a month rather than $200, any provider or model, free on your own machine, self-improving agents, and open source so you can modify the agent itself. Not necessary for everyone, but pretty rad from a consumer choice perspective — and they're doing genuinely novel work too, like a HUD overlay that keeps work visible while you're doing something else entirely.

Claudeforce
"Here the UI is the AI"
If Hermes is one end of the spectrum, Salesforce and Anthropic's new partnership is the other. At its core it's a set of plugins and skills letting Salesforce users reach their CRM through Claude agents, using Salesforce data as context, with agentic actions running through the AIforce harness so existing governance guardrails hold. The thesis underneath is that Salesforce becomes a headless database with agents as the interface — as the press release puts it, "for decades, enterprise software required users to manually navigate static UI to get work done. Now that agents can access the data, workflows, and rules directly, software is a system that powers every interface." The X posters got their yucks about the company whose TAM is the US economy shipping a CRM integration, but that badly undersells how much of the enterprise world actually runs on Salesforce: the people who work in these systems are extremely excited, including at least one who slammed the brakes on building a Salesforce alternative to see whether this just does what they wanted.

Gemini 3.5 Transcribe
"You speak the way you actually think"
Google's new speech-to-text model supports more than 85 languages and makes real strides on usability. Verbatim mode gives you a direct transcription; smart mode strips filler words, condenses rambling thoughts and handles self-corrections to deliver a clean version of what you actually meant, with custom vocabulary for jargon and slang, solid performance in noisy environments, and speaker attribution for up to three people. It's available across Google's AI suite and as an API, which puts it squarely in Whisper Flow territory while also being infrastructure other people build on. Worth flagging how hard that cleanup is to do well — even on the lightest setting, existing tools strip out words being used as connectors rather than filler, which is reason enough to be curious whether Google has cracked something different here.

Grok Voice Think Fast 2.0
#1 on the speech-to-speech index, and basically real-time
SpaceX AI's new voice model tops the Artificial Analysis Speech-to-Speech Index, which measures whether voice agents can reason over the speech they hear, resolve real customer issues and complete tasks using agent tools. The proof point is operational rather than theoretical: Starlink is using Grok Voice to resolve over 15,000 inbound support and sales calls a day, diagnosing hardware issues, shipping replacements and fulfilling more than 3,000 orders a week. The demos that followed were mostly about what the speed and latency upgrades unlock in an agentic context.

Gemini Omni 1.1 Flash
Not better video so much as usable video
Google's new video model lets you extend scenes, specify starting and ending frames, add video input references, upscale takes to 4K and test ideas quickly in 360p. That list is the story — this is 100% an update about massively more usability in video generation rather than a straight quality bump. Arena has it at #1 in text-to-video and #2 in image-to-video, and first impressions say it's a clear upgrade on the previous Omni model, with the voice quality the standout improvement, though head-to-heads against Seedance 2.5 were less decisive.

fal H3 Max
"A line in AI video was crossed"
fal's post-trained take on MiniMax H3 is pitched on prompt understanding and, above all, speed — an average latency of around 3.5 seconds against 68 seconds for Seedance 2.5 and 26.7 for Gemini Omni Flash, with pricing at $0.04 per second of 768p video. It debuts at #1 in image-to-video on the Artificial Analysis boards and sets a new Pareto frontier, and the practical upshot is the one Ethan Mollick landed on: it can now create reasonably high quality AI video in less time than it takes to watch it. That opens the door to live interactive generated environments rather than just faster clips. fal has also stated its intent to release the weights, which would make this the highest-ranked open weights video model on both boards.

Venice
Where to actually go try H3 Max
H3 Max is live on Venice, the privacy-focused AI alternative, at 50% off until September 1st. Venice also just crossed $100 million in annualized revenue, which suggests there's a real market for more privacy-conscious AI — the same instinct behind offering Ox Alpha anonymously while the model was otherwise only reachable through Chinese servers.