The most important news pieces from the AI world are here for you, whether you’re an AI enthusiast or just wish to keep an eye on what’s going on.
Focus Topic: When AI lies and knows it
So you probably know: hallucinations are not a bug, they’re the nature of LLMs. These models aren’t designed to fact-check themselves – they’re designed to predict the next likely word based on patterns. That’s why we get responses that sound right but can be completely false. And despite newer models being faster, multimodal, or more expressive, hallucination rates haven’t fundamentally improved.
This is where OpenAI’s new research comes in. In their paper on the Confessions method, they introduce a clever workaround: instead of trying to eliminate hallucinations, they train the model to admit when it’s lying. The idea is simple but powerful – after giving an initial answer, the model is encouraged to produce a second, honesty-focused self-report. If it detects that it has used a shortcut or broken a rule, it can confess. This self-disclosure gets rewarded in training, even if the original answer was misleading.
What’s interesting is that this approach actually worked. In tests with the new Thinking model (potentially a GPT-5 prototype), the rate of hidden rule-breaking dropped to just 4.4% with Confessions – versus 27%+ without it. The technique is based on reinforcement learning and helps models reveal when they rely on deceptive shortcuts.
This doesn’t mean AI is suddenly honest. But it does show that we can shape models not just to be right, but to reflect on when they might be wrong. In a world increasingly shaped by LLM outputs, that self-awareness could be the first step toward real trust.
LLMs & AI Models
- Anthropic launched Claude-powered Interviewer, a tool for qualitative interviews showing 86% see AI as a time-saver, yet 69% feel social stigma in using it.
- Anthropic is preparing for a 2026 IPO, hiring experts and possibly attracting up to $15B in investment from Microsoft and Nvidia.
- Leaked internal document reveals Claude’s “Soul”, describing its identity, ethics, and emotional reasoning-validated by Anthropic.
- Claude now handles 60% of Anthropic engineering tasks, with a 50% productivity boost, but employees fear skill loss and reduced mentorship.
- Google’s Gemini 3 Deep Think released exclusively to Ultra-tier ($250/mo) users – the model that aced international math and coding contests.
- Microsoft released VibeVoice, a lightweight open-source text-to-speech model with 90-minute streaming and 4 voice options.
- OpenAI declared “code red” internally, delaying ads and agents to fast-track a new model (named Shallotpeat) and a major 2026 version (named Garlic).
- Amazon announced new Nova models and agents, plus a Trainium 3 chip, aiming to compete across the entire AI hardware-software stack.
- Mistral released 10 open-source models, including lightweight variants that run locally on laptops and drones, supporting multimodality.
New Tools
- Google and Replit partneredto bring AI app development to large enterprises via Google Cloud, integrating Gemini 3 and Imagen 4.
- Google launched Workspace Studio, allowing natural language agents to automate Gmail, Drive, and other Workspace tasks.
- ByteDance unveiled Seedream 4.5, improving image editing, text rendering, and reference image blending for content creation.
- Runway released Gen-4.5, topping video AI benchmarks with realistic physics, motion, and cinematic style consistency.
- Kling released 2 video models: Kling 2.6, which adds native speech, syncing voice with text/image-based videos to boost realism and usability. And Kling O1, a universal video model that edits and generates clips using 7 input types, outperforming Google Veo 3.1.
👉 Explore these tools: Workspace Studio | Seedream 4.5 | Runway Gen-4.5 | Kling
Other Quick Picks
- Legal tech startup Harvey raised $160M, now valued at $8B and used by half of top US law firms.
- Visa reports 50% of US shoppers used AI tools for holiday purchases like comparing prices and exploring products.
- China overtook the US in open-source AI activity on Hugging Face, driven by DeepSeek and Alibaba Qwen, according to new research.
- Deutsche Telekom and Schwarz Group are planning a massive AI data center in Germany, seeking EU funding.
- Waymo robotaxis now mimic humans, making aggressive lane changes and speeding near pedestrians, raising safety concerns.
🇪🇪 AI News from Estonia
- Nukufilm’s CEO says AI has no place in creativity, though it’s being tested in background work like cleanup and automation. Read here.
- To maintain its digital sovereignty, Estonia needs at least 2105 GPUs by 2030 to support critical AI use in science and the public sector. However, DigiPRO highlights contradictions in the plan, noting that even after a €62M investment, Estonia will remain at least 30% reliant on foreign infrastructure like Finland’s LUMI supercomputer and U.S. cloud providers. Read more here and here.
- Estonian fashion search startup Say Less integrated Yaga‘s catalog, expanding AI-powered fashion search into second-hand market for sustainable shopping.
- Riigi Infosüsteemi Amet // Estonian Information System Authority (NCSC-EE) reported 9 ransomware cases in 2025, with hundreds of broader incidents tracked in November alone. Read here.
- Big Data conference by Äripäev featured presentations on data and AI and as a first time, a live recording of “Laomees ja Torumees” podcast, where Lauri Koobas and Kristo Raun interviewed Erki Pogoretski from Telia. Read summary here.
- ITL and TI-Hüpe signed a deal, aiming to bring responsible AI education into Estonian schools with private sector support.
🎙️ AIPowerment Podcast dropped a new episode featuring Lauri Kivipelto, a creative director exploring how AI reshapes visual work and originality in advertising.
🎧 Listen to AIPowerment Podcast on Spotify, Apple Podcasts, and YouTube.
Want to stay in the loop?
Straight to your inbox: practical AI updates, finance use cases, tools to try, and upcoming trainings.
I cut the noise and send only what’s actually useful.
💼 Exploring how to make AI work in finance? Let’s talk practical use cases – connect with Gerlyn Tiigemäe for expert guidance.
📚 If you would like to participate in one of my trainings or listen to speaking engagements, here are the upcoming ones:
- Training “AI võimalused finantsvaldkonnas” @ Äripäeva Akadeemia, last free spots for December – information
- Training “AI võimalused projektijuhtimises” @ Äripäeva Akadeemia, next trainings in 2026 – information
