How much stock should you order for next week? For most businesses, that's still part guesswork. Mistral has released Timestral, a family of forecasting models, and it's aimed at exactly that kind of question. Timestral predicts what comes next in numbers that change over time, like daily sales, energy use or deliveries. Instead of one guess, it gives a range, so you can see how confident the forecast is. It can also use things you already know about the future, like holidays, planned shutdowns or weather forecasts. It comes in 4 sizes: the smaller ones are faster and cheaper to run, and the bigger ones are more accurate.
Why it matters to you
Ordering, staffing and cash planning all start with a forecast, so a better one means fewer empty shelves and less waste. Mistral says it doesn't need a graphics chip, and that the smallest version forecasts a day ahead in about 10 milliseconds on a standard server. Mistral's own example is a factory that uses it to catch equipment problems before they turn into breakdowns. Picture a bakery planning a public holiday weekend: the forecast already knows it's a holiday, and shows a low and a high case.
The risks
The headline results come from public benchmark tests, which means standard test datasets, and your own sales data can behave very differently. A forecast is not a promise, so look at the whole range it gives you, not just the middle number. And it needs someone comfortable with data and code to set up. It isn't an app you just download.
Try it safely
Ask yourself: which number would you most like to know a week early? Sales, stock or cash? Start there. Mistral has published the code on GitHub so you can run it on your own data, under what it calls a modified MIT licence. If you have a year or 2 of daily sales, ask whoever handles your data to test it against what actually happened last month.
Annie's take This is the least flashy story today and maybe the most useful, because forecasting touches ordering, staff and cash. If your records are in good shape, it's worth a test.
What if your AI assistant didn't wait for you to type, and instead messaged you each morning with what needs your attention? That's the idea behind Claude Duet, which its makers have just shown as a research preview. Claude Duet is an agent, which means software that carries out tasks for you, not just answers questions. According to its product page, it sends you a morning brief, drafts next steps, and keeps working while you're away. It stops and asks for your okay before sensitive steps, like sending a message or making a payment. Think of it less like a chatbot and more like an assistant who checks in with you.
Why it matters to you
For a small business owner, the examples will feel familiar: chasing unpaid invoices, preparing payroll for review, and getting contracts ready to send. In the product's own example, the morning brief flags a hiring deadline, a contract change and a permit that needs approval. Picture a café owner who wakes up to one short list: what needs her today, with the replies already drafted for her to approve. It works with more than 6,000 connectors and plugins, so in theory it can pull your calendar, inbox and sales tools into one place.
The risks
Its makers say it's a research preview, and that it can misread a thread or get wrong what's urgent. It also means giving an AI access to your inbox, your calendar and possibly payments, so start with low-stakes jobs and keep the approvals on. And it shares your plan's usage limits, so big jobs like a slide deck or a financial model will use up more of them.
Try it safely
Ask yourself: which 3 admin jobs eat your mornings? Those are the ones to hand over, with approvals on. Right now there's a waitlist for Pro and Max subscribers in the US, so most of us can only watch for now. If you get in, look for 3 things: does the brief match what you'd have flagged, are the drafts usable, and does it always ask before acting.
Annie's take This is where AI assistants are heading, because waiting for you to type a prompt is the slow part. But I'd treat it like a new hire on probation: useful, watched closely, and not handed the company card on day 1.
There's a new place to put your business in front of customers: right next to ChatGPT's answers. OpenAI says ChatGPT Ads now runs in 26 more countries, and that over 100,000 advertisers have used it. In plain words, when people use ChatGPT to explore options and decide what to get, businesses can pay to show up there as an ad. OpenAI says the ads are kept separate from the answers, and are matched to what the person is trying to do in the conversation. It's aimed at everyone from local businesses to global brands, according to OpenAI. Think of it like search ads, except the question comes up in a conversation instead of a search box.
Why it matters to you
If your customers start asking a chatbot which accountant or coffee grinder to choose, instead of searching, this is where they'll see offers. For a small business, that's a new ad channel to watch, next to search ads and social media ads. Picture a florist: someone asks for anniversary gift ideas, and her shop's ad appears beside the answer.
The risks
Today's sources give no prices and no results data, so treat any early success stories you hear with care. Some people may not like ads in a chat, so watch how your own customers react before you spend much. And because it's new, the rules, formats and countries may keep changing, so check again before you plan a budget.
Try it safely
Ask yourself: do your customers ask questions before they choose you? If they do, this channel is worth watching. If you want to test it, start with a small budget and one clear offer, and compare it with what you already pay for search ads. Look for 3 things: whether your country is now included, what you can target, and how you'll measure sales, not just clicks.
Annie's take Worth a small test if your customers ask a chatbot before they spend, because it's easier to learn a new channel while it's young. If most of your customers walk in off the street, I'd wait and watch.
Imagine calling a supplier abroad, speaking your own language, and hearing them reply in yours, while you talk. ElevenLabs has launched Eleven Realtime Translation, a live voice translation model, as an early alpha for its enterprise customers. Each person on the call speaks their own language and hears the other side in theirs, in a natural voice, during the conversation. The company says it covers more than 90 languages, spots the language by itself, and keeps one consistent voice for the whole call. Teams can choose between better translation for tricky calls, like billing or claims, or the fastest back and forth. You can pick any voice from the ElevenLabs library, or a clone of your own brand voice, and it stays the same for the whole call.
Why it matters to you
For a business, the obvious use is customer support: one support person could help callers in many more languages without an interpreter on the line. Picture a small travel agency with 3 staff, taking calls from customers in several countries. ElevenLabs also describes phone networks building it into calls, and live events where everyone hears the speaker in their own language. It says Deutsche Telekom is already using it to build new calling features.
The risks
It's an alpha, which means early and unfinished, and it's only for enterprise customers through the API, the connection developers build on. A wrong word on a billing call could cost you, so keep a person checking anything important. And live customer calls would pass through another company's system, so check what your customers have agreed to.
Try it safely
Ask yourself: how many customers do you lose today because nobody on your team speaks their language? If you run a support team with callers in several languages, it's worth asking ElevenLabs about the alpha. When you test it, use real calls and check names, numbers and prices in both languages, because small slips matter most there.
Annie's take Live translation could quietly change who a small team can serve, because language stops being the limit. But I'd wait until it's out of alpha before using it on any call that involves money.
Your next work laptop might run AI on the machine itself, instead of sending everything to the cloud. NVIDIA says its RTX Spark laptops are now available, from ASUS, Dell, HP, Lenovo, Microsoft and MSI. They're Windows laptops with an NVIDIA chip that combines graphics and processing, and up to 128 gigabytes of memory. NVIDIA says that's up to 1 petaflop of AI computing, which means a huge amount of number crunching for something you can carry. Prices start at 2,399 US dollars, according to NVIDIA. Laptops from those makers are available now, and NVIDIA says more systems are on the way.
Why it matters to you
Running AI on the laptop itself, what NVIDIA calls local AI, can mean your files don't have to leave the machine. Picture a small video team: editing, design work and AI tools on one portable machine, without waiting on an online service. NVIDIA also publishes guides and recipes for running local AI on these machines. For teams that handle private client files, that can make AI easier to say yes to.
The risks
These are premium machines, and that's a starting price, so check how much memory you actually get for it. Most AI tools your team uses today run online anyway, so a powerful laptop won't change those. And software that makes full use of these laptops will take time to arrive.
Try it safely
Ask yourself: what would you run on the laptop that you can't run online today? If the answer is nothing, keep your current machine. If you're replacing laptops soon, list the AI tools your team would actually run on the machine before paying for the extra power. When you compare models, look at memory, because it decides how big an AI model the laptop can run.
Annie's take Good news for creators and teams with private files, because more of their AI work can stay on their own machine. For everyone else, it's a nice to have this year, not a must have.