- Synthetic
- Posts
- AI Hallucinations Are Getting Worse đ”âđ«
AI Hallucinations Are Getting Worse đ”âđ«
Plus, Road Rage Victim 'Speaks' At His Killer's Sentencing đšđ»ââïž
Subscribe to Synthetic
This weekâs most interesting and relevant AI news and analysis
Remember, you can read past editions of Synthetic here.
This Week in AI

AI Hallucinations are getting worse, not better
The Smarter AI Gets, The More It Hallucinates đ”âđ«
Since the development of the very first large language models, AI has hallucinated, confidently making statements that arenât accurate. Why AI models do this is debated, but they tend to fabricate responses when they arenât properly grounded in fact. Also, large language models donât fundamentally understand the difference between answering a factual question and writing a poem about a little green frog. Sometimes they get âcreativeâ when they shouldnât. Researchers hoped that larger models, trained on more data, and imbued with reasoning capabilities would be more accurate than the previous round of models. The opposite has happened. OpenAIâs latest o3 and o4-mini reasoning models were found to hallucinate more than earlier versions. Their in-house benchmark found that o4-mini hallucinates 48% of the time, and o3 does it 33% of the time.
Syntheticâs Take: Accuracy is a key issue for enterprise AI applications. Companies wonât deploy AI that feeds customers incorrect information. Grounding models in company data using vector databases and RAG (or soon, MCP) helps. But, accuracy was supposed to improve, not diminish, with the latest generation of models. There are two big implications: 1) The promised era of agentic AI may be delayed, and 2) Yann LeCun is likely correct when he says that language models are not the path to human-level intelligence, so-called AGI.
Speaking at the Milken Institute Conference this week, Nvidia CEO Jensen Huang was direct with the audience about the impact of AI: âYou will not lose your job to AI, but will lose it to someone who uses it. I recommend 100% take advantage of AI, donât be that person.â View the whole conversation here (29 mins).
In a moment that feels more like a scene from Black Mirror, family members of an Arizona man killed in 2021 during a road rage incident played an AI-generated video of the victim, where he gave his own victim statement and spoke about his faith and forgiveness. A video of the AI-generated statement is included in the article.
Video: NVIDIA CEO Jensen Huang Speaking at the 2025 Hill and Valley Forum
Each year, the Hill and Valley Forum brings together people from the US government and Silicon Valley (Hill and Valley, get it?). In this video, Jensen Huang, wearing a dapper suit rather than his usual trademark leather jacket, discusses AI factories, the three layers of AI, physical AI, how AI will revolutionize every industry, and the impact AI will have on jobs. ââNew jobs will be created. Some jobs will be lost. Every job will be changed.â (24 mins)
âFifty percent of the worldâs AI researchers are Chineseâ
AI Tech and Innovation

Will you empower your AI agent to shop for you?
AI agents will transform digital commerce and the online experience. People are already using AI tools like Perplexity to search for and discover products, but itâs still difficult (and inadvisable) to have AI purchase things for you. Visa worked with OpenAI, Anthropic, Microsoft, Mistral, Stripe, IBM, and others to build a secure payment system for AI agents. Consumers can give AI agents spending limits and conditions on purchasing items. đïž
Gathering at the Robotics Summit & Expo in Boston, robotics leaders are all worried about the impact tariffs will have on their growing industry. Robots need sensors, semiconductors, rare earth magnets, and a large battery to power their anatomy. Many of these components come from China. Tesla CEO Elon Musk warned investors that a shortage of rare earth magnets will delay the development of the Optimus humanoid robot. Some business leaders expect tariffs to accelerate the demand for robotics as companies re-shore operations and look to automation to deal with high labor costs. đ€
Apple plans to turbocharge its popular Xcode developer platform with AI. The new version of Xcode will use Anthropicâs Claude Sonnet model to help coders write, edit, and debug software. Initially, the latest Xcode will only be used internally, and Apple has yet to announce a public release. In related news, OpenAI will acquire Windsurf (formerly known as Codium), a leading AI-powered coding platform, for about $3 billion.
Syntheticâs Take: The first programmers had to program in machine codeâbinary or hexadecimal numbers. Assembly language made it a little easier. Then came low-level languages like Fortran, COBOL, and BASIC. Object-oriented languages like C++ and versatile languages like Python have made it easier to turn ideas into software. Each time, we abstract away from the underlying machine code and make programming more accessible. AI will soon become the coding paradigm that all coders use, making it easier to write, edit, test, and debug software, in collaboration with an AI coding assistant.
AI Insights

The iPhone as we know it today may be gone within a decade, replaced by AI
While giving testimony at the antitrust case against Google, Eddy Cue, Apple Senior Vice President of Services, said that AI may eventually replace the iPhone. Cue oversees Appleâs massive services portfolio, which includes Music, News, Maps, Podcasts, Apple TV+, Apple Pay, Ads, Apple iCloud, and all of Appleâs creativity and productivity apps from Mail to Final Cut Pro. He was behind the creation of the iTunes Store and the App Store. Cue said, âYou may not need an iPhone 10 years from now, as crazy as it sounds. The only way you truly have true competition is when you have technology shiftsâŠAI is a new technology shift, and itâs creating new opportunities for new entrants.â Cue added that Apple is âactively looking atâ moving the focus of its Safari web browser from search to AI-based search.
Syntheticâs Take: AI is shaking things up for Google and Apple. AI-powered search is usurping the ten blue links of Google search. Apple may soon remove Google as the primary search tool on its devices. With time, AI assistants will make apps less relevant and eventually replace them altogether. Why do you need an Uber app when your AI assistant summons a car for you? Apps become API calls or tools for agents. Appleâs reworking of its notoriously lame Siri assistant has been slow and fraught with delays and internal conflict. Itâs been a rough ride for Google, too, with DOJ scrutiny and a potential break-up of its businesses on the horizon. This has been reflected in significant downward pressure on Alphabet stock, even though Googleâs businesses are likely worth more apart than unified under the Google umbrella.
JPMorgan has long been known as a leading adopter of AI. It has identified 450 potential AI use cases, and CEO Jamie Dimon expects that number to surge beyond 1,000 by next year. This week, JPMorgan Asset Management CEO Mary Callahan Erdoes said the company increased gross sales 20% using AI between 2023 and 2024. JPMorgan's call volume skyrocketed in April as tariff news rocked global financial markets and panicked investors. Private client advisors use an internal AI tool, Coach AI, to find content and research up to 95% faster so they can focus on client conversations and minimize the time spent searching for information.
As reported last week in Synthetic, language company Duolingo has stated it will become an âAI-firstâ company and begin to replace contractors with AI. They are already using generative AI to create content. In a press release, Duolingo CEO Luis von Ahn said, âDeveloping our first 100 courses took about 12 years, and now, in about a year, weâre able to create and launch nearly 150 new coursesâŠ.This launch reflects the incredible impact of our AI and automation investments, which have allowed us to scale at unprecedented speed and quality.â đŠ
Toolkit for the Future
Building a helpful AI assistant for your website is so easy now. Just pick a hyper-realistic voice, give your new assistant a simple prompt (e.g., âYou are a helpful customer service agent for a keynote speakerâ), upload relevant FAQ documents, and integrate it with your website. Done! We used ElevenLabs for stevebrown.ai, and love it. Super easy to set up and configure. ElevenLabs is also great for creating audiobooks or multilingual voice tracks in up to 29 languages to reach new markets. Try it for free. Itâs one of the best AI solutions out there.
Create and sell online courses from your website with this top-rated platform. Itâs easy to use with hundreds of beautiful templates and a powerful generative AI assistant that makes content generation a breeze. This is the platform Synthetic has selected to launch our new AI training course, coming this summer. The AI helps you create lessons, exams, assessments, and more. Free to try. Use promo code 10XR to get 30% off your first two months.
With a support automation platform, you can relieve pressure on your support team and serve customers better and faster. AI answers 90% of queries and escalates the rest to your team. It automatically updates its knowledge database so itâll never escalate the same issue again. The platform supports SMS, social, email, and phone communication.
Focus on the email that matters. Deep Clean automatically identifies old, unimportant emails so you can delete them. Snooze emails until a better time and easily track emails that need a response. Automatically filter unimportant emails into specific folders. It works across every device where you check your email. Get to inbox zero with award-winning SaneBox.
Try this easy-to-use, AI-powered alternative to DocuSign to write proposals, create quotes, get eSignatures on contracts, build online forms, collect payments, and much more. Great for sales, HR, legal, finance, and IT. Free 14-day trial.