Ben's Bites

Where can my team start using AI? This is a common question in our community. Do I hire a consultant? Hire for a new AI expert role? The simple way to start is with people already on your payroll. Try this: Let your team explore our tutorial library for an hour. We've organized it by function - marketing, sales, data, and operations. Ask them to pick 4-5 tutorials that resonate with their daily work—the tasks they spend hours on each week. You can start with our free tutorials to test the waters. For the next week, let them experiment with their list and note down other ideas where they think AI could help. Ask them for another list of 5-10 highly specific use cases. This creates natural buy-in (they choose what to try) and builds practical skills through real work. Once you see the impact, our pro tutorials (we have team plans) can help scale these wins across your company—still cheaper than “AI Consultants”. More importantly, you'll discover patterns where standardizing AI use could benefit the entire company. Maybe multiple teams need help with email drafts, or everyone's trying to analyze spreadsheets. These patterns help you focus on high-impact areas first and guide you on where to invest time and money. Ready to get your team started? Check out our tutorial catalog: https://lnkd.in/dGFFpUpg

FLUX models by Black Forest Labs have become the go-to choice for image-generation apps Their FLUX1.1 [pro] now has two new modes: Ultra can pump out 4MP images (that's 4x higher res!) with the same speed, while Raw mode makes humans and nature scenes look way more natural and less "AI-ish" More info: https://lnkd.in/d7ntw6ev

chat.com now points to ChatGPT Dharmesh Shah, CTO of HubSpot, who bought the domain last year for (reportedly) $15.5M, gave a cryptic prompt to figure out the sale price

Claude 3.5 Haiku is here (only for developers), but we’re not impressed Overall it feels weaker than GPT-4o-mini and Gemini 1.5 Flash 002 while being much more expensive The redeemable quality of this model can be its coding capacity, where Anthropic claims it’s even better than the old Claude 3.5 Sonnet ...but it’s not the reason to stop loving Claude. It just got two game-changer features You can now dictate up to 10 minutes of voice messages on the Claude mobile app and Claude now supports PDF uploads with images and charts in it too (up to 100 pages) Haiku announcement: https://lnkd.in/dCx_vPdt

ChatGPT Search is live for Plus and Team users After running the SearchGPT prototype, OpenAI has finally added it to the main ChatGPT experience. It searches the web automatically when it thinks it might be necessary (which might be annoying for some) This is OpenAI's biggest move yet to compete with Google. Instead of multiple searches and link-digging, users can have natural conversations about current events, weather, stocks, sports and more—with ChatGPT automatically pulling relevant web sources. They've also partnered with major publishers like AP, Reuters, FT and others to ensure quality information. Some practical ways to use it at work: - Writing reports? Get the latest industry updates without leaving your draft - Checking competition? Ask about their recent product launches or news - Need market stats? Pull current data while working on your deck Our verdict: if you're tired of switching between Google and your work, ChatGPT search makes life easier. But I wouldn't unpin Google from the browser just yet.

OpenAI released a new benchmark called Simple QA. It includes 4000+ single-line factual questions. It’s designed to be hard for most large language models. Claude and GPT models perform similarly on the benchmark but numbers for Gemini were not reported. We ran 10% of those questions with Gemini models. The score for Gemini-1.5-Flash is comparable to GPT-4o-mini and the same is true for Gemini-1.5-Pro and GPT-4o. But, let’s get a little geeky now. The official benchmark uses GPT-4o (a large model) to grade the answers. We used Gemini Flash, a much smaller and cheaper model. We scanned through a 100s of results and we didn’t find any errors in Flash’s grading. It’s an important distinction on how to use these models i.e. given all the important information, Flash’s reasoning is very strong. But if you leave the job of determining facts to it, you are going to get the wrong results. We tested roughly a dozen of the questions that Gemini 1.5 Pro got wrong on the Gemini web app. Since it has access to web searches, we got accurate answers to all of them (even in the free version of Gemini). Google is also adding this web search capability to Gemini API and Google AI studio (starting later today) to ground the model’s answers with Google search. Read more stories on the Ben’s Bites blog: https://lnkd.in/dz3Ehn5i

Top new AI tools discovered this week

GitHub's AI strategy is expanding beyond OpenAI Copilot now supports both Anthropic's Claude and Google's Gemini models alongside GPT Plus, GitHub just launched Spark - a new tool that lets anyone create micro-apps using natural language: https://lnkd.in/dp9Ra_jD

We're hiring! We're on the lookout for freelance creators that can make educational content for our AI catalog. - Create 4-7 bitesized tutorials per week - Work from anywhere - We pay per tutorial so the more you create, the more you get paid - $125 USD per tutorial - We also cover the subscription cost for to use any specialist AI tools Full details and application form here: https://tally.so/r/mDJgDp

ChatGPT’s web app now has an option to search old chats. A much-needed feature All chats (and the content inside, not just titles) are searchable Also, both the MacOS and Windows desktop apps for ChatGPT have access to Advanced Voice Mode now More info on this thread: https://lnkd.in/d57vENZ4