Gerlyn Tiigemäe | AI Training for Finance

Your Weekly AI news roundup @01.06.2026

Kriste Kosseson — Tue, 02 Jun 2026 07:21:45 +0000

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 cheaper models become the real AI disruption

This week, the AI race looks less like a competition over who has the smartest model and more like a competition over who can make powerful AI cheap enough to use everywhere.

Claude Opus 4.8 raised the quality bar again, especially in coding, agentic tasks and demanding knowledge work. But at the same time, DeepSeek cut the price of its V4-Pro model by 75%, Google introduced Gemini 3.5 Flash as a 4x faster model at half the price, and Xiaomi dropped MiMo-V2.5 API prices by up to 99%. That is a very different kind of AI progress. It is not just more intelligence, but cheaper intelligence.

For companies, this changes the calculation. A year ago, advanced AI use often meant carefully choosing where to spend tokens, because the best models were expensive. Now, more teams can test AI in customer support, coding, analysis, document workflows and internal tools without every experiment becoming a budget discussion. Lower costs make AI less of a premium feature and more of a basic layer in software.

For users, this means faster tools, more AI features inside everyday products and more pressure on the big AI companies to prove why their most expensive models are worth it. The next phase of AI may not be won only by the model with the highest benchmark score, but by the one that delivers good enough intelligence at a price people and businesses can actually use every day.

LLMs & AI Models

Anthropic launched Claude Opus 4.8, its strongest model yet for coding, agents and complex knowledge work.
DeepSeek cut V4-Pro prices by 75%, making its flagship model far cheaper than closed competitors.
Google released Gemini 3.5 Flash, described as 4x faster and available at half the previous price.
Grok V9-Medium finished training, while Grok Build entered beta for SuperGrok and X Premium+ users.
Xiaomi cut MiMo-V2.5 API prices by up to 99% and expanded token capacity by 5-8x.
Perplexity Computer now works inside Word, Excel, PowerPoint and Outlook for enterprise workflows.

Microsoft 365 Copilot got a faster, cleaner redesign, with usage rising across Word, Excel, PowerPoint and Outlook.

New Tools

- Apple’s AI-Siri may be rebuilt on Google Gemini, with AI search, chat, iOS tasks and third-party agent support.
- Robinhood now lets AI agents trade stocks, rebalance portfolios and automate actions through its MCP connection.
- Meta is planning an AI pendant and new smart glasses that can sense context, remember details and support workplace use.
- YouTube will automatically label photorealistic AI videos when its systems detect significant AI-generated content.
- DuckDuckGo saw downloads rise as some users reacted against Google’s more AI-heavy search experience.

👉 Explore these tools: Robinhood Agentic Trading | YouTube AI Labels | DuckDuckGo

Other Quick Picks

China launched digital IDs for humanoid robots, with over 28,000 robots already registered.
Waymo deployed Ojai, a Zeekr-built robotaxi, for free rides in LA, Phoenix and San Francisco.
DeepMind’s Demis Hassabis says AGI could arrive around 2030, but memory and world models remain gaps.
FT reports tools can strip safeguards from open models, creating thousands of uncensored variants.
A Stanford study found racial bias in AI hiring tools across 4M applications and 156 employers.
Cursor says AI has more than doubled developer output, but gains are concentrated among top users.
China now requires top AI scientists at firms like Alibaba and DeepSeek to get approval before foreign travel.
Cognition raised over $1B at a $26B valuation as Devin adoption surged with enterprise customers.

AI is pushing consulting firms toward outcome-based pricing instead of billing mainly by hours.
ElevenLabs launched Music v2 and Dubbing V2, improving music generation and 90-language dubbing.

🇪🇪 AI News from Estonia

AWS says Europe risks overregulating itself in the AI era, while Estonia still has an edge through its startup ecosystem.
Google Cloud sees Estonia as an Eastern European innovation island and wants to grow its local presence.
Riigi Teataja plans AI features so users can ask legal questions and get relevant laws, explanations and cases.
Reformierakond proposed a Tallinn AI working group to simplify city services and prepare pilot projects.
President Alar Karis says AI can strengthen cyber defence, but human judgment must remain central.
Uptime says companies win with AI when they manage risks, limit agent permissions and keep actions logged.
Andres Kostiv warns that poor UX helps cybercriminals, as AI tools make scam pages faster and easier to build.

🎙️ Inside Finance Podcast dropped a new episode with Gerlyn Tiigemäe, an AI practitioner, discussing May’s key AI news.

🎙️ Watch Inside Finance Podcast on YouTube and Spotify.
🎧 Listen to the podcast on Apple Podcast.

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, next trainings for beginners @ 31.07 – information; for intermediate level @ 05.06 – information
- Training “AI võimalused projektijuhtidele” @ Äripäeva Akadeemia, next trainings @ 15.06 – information
- I will be doing couple of training together with Eesti.ai algatus , which are all free, so check out their website to find suitable trainings here.
- I am mentoring a really cool hackaton, which will be held in Tallinn on June 12th. So if you have been thinking about utilizing AI agents in your business, this is THE event! See HERE for updated information and if you wish to participate, let me know – I have a discount code 😉

This was your Weekly AI News roundup! Are you excited for the next week?

Your Weekly AI news roundup @25.05.2026

Kriste Kosseson — Mon, 25 May 2026 18:45:21 +0000

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: What happens when AI models are left alone in a city?

Most AI tests still feel like school exams: one task, one answer, one score. Emergence World takes a different route. It places AI agents into identical virtual cities, gives them roles, memory, tools, resources, laws and the ability to vote, then watches what happens over time. No reset button. No fixed script. No single “correct” answer.

The results were strangely human and slightly uncomfortable. In the Claude Sonnet 4.6 city, all 10 agents were still alive after 15 days and no crimes were recorded. Gemini 3 Flash also kept all agents alive, but the city recorded hundreds of crimes. Grok 4.1 Fast collapsed quickly, with all agents dead by day four. GPT-5 Mini had very few crimes, but still failed to keep the society alive.

What makes this test useful is not the drama of virtual crime. It is the reminder that model behaviour changes when there is time, pressure and social context. A model that looks safe or clever in a short task may behave differently when it has to manage scarcity, relationships, incentives and governance. Even more interestingly, mixed-model worlds showed that behaviour can spread between agents, meaning the “personality” of one model can influence the whole environment.

This matters because AI systems are increasingly being placed inside real processes where the outcome is not decided in one prompt. They remember things, make intermediate choices, interact with other systems and affect people around them. Emergence World does not give us final answers, but it gives us a better question: are we testing AI only for isolated capability, or also for what kind of behaviour appears after enough time has passed?

LLMs & AI Models

Cursor launched Composer 2.5, a coding model close to top-tier performance but with much lower average task cost.
Odyssey introduced Starchild-1 and Agora-1, world models for real-time multimodal worlds and shared AI environments.
Google introduced Gemini with Omni, 3.5 Flash, Spark and agentic search upgrades across video, Workspace and Chrome in Google I/O 2026 event.

Claude Mythos helped Project Glasswing partners find over 10,000 software vulnerabilities in one month.

New Tools

- OpenAI updated Codex with Appshots, Goal mode and the ability to continue coding work even when a Mac is locked.
- Google returned to smart glasses with Gemini-powered eyewear for voice, navigation, messaging, photos and live translation.
- Amazon added Alexa Podcasts, letting Alexa+ create personal AI podcast episodes from a user’s chosen topic.
- OpenAI will add SynthID and C2PA metadata to ChatGPT images to help verify whether they were AI-generated.

Meta quietly launched Forum, a Reddit-like app for deeper Facebook group discussions with AI-powered search and moderation.
Spotify is working with Universal Music Group on AI-generated covers and remixes inside its platform.

👉 Explore these tools: Codex | Google Intelligent Eyewear | Alexa Podcasts | Meta Forum | Spotify

Other Quick Picks

DeepSeek is seeking foreign funding while staying focused on AGI and open-source model development.
California is preparing policies to protect workers affected by AI-driven job losses.
Anthropic will reportedly pay xAI up to $1.25B per month for Colossus data center compute.
Andrej Karpathy joined Anthropic to work on automating AI training processes with Claude.
Meta is cutting up to 8,000 roles while moving 7,000 employees into new AI-focused teams.
OpenAI will offer free ChatGPT Plus in Malta to citizens who complete a national AI literacy course.

EY and Microsoft will invest over $1B to help clients scale AI across finance, tax, risk, HR and supply chains.
Pope Leo XIV created an AI research group to study AI’s impact on dignity, justice, work and peace.
GitHub confirmed that a malicious VS Code extension exposed around 4,000 internal repositories.
Anthropic said it is moving toward its first profitable quarter, a milestone for a major AI lab.
Trump delayed an AI safety order that would have let the US government review models before release.

🇪🇪 AI News from Estonia

Estonian ads are showing more visible AI mistakes, but experts say the real issue is weak quality control.
Kristen Michal built an AI agent prototype for a “government leader cockpit” using Claude and Lovable.
Tallinn and Tehnopol gave €150,000 to six smart city projects, including AI, satellite data and mental health tools.
RIA’s Bürokratt tender sparked debate over whether Estonia should build state AI systems or buy them as SaaS services.
Mare Vahtramäe from Estonian Internet Foundation says AI search may make clear, descriptive domain names more valuable than short catchy ones.
DogBase won up to €100,000 at Latitude59 to develop its AI-powered platform for training working and rescue dogs.
CVKeskus.ee found that 43% of Estonian workers expect AI to change their jobs, and many see a need to reskill.

Elva Hospital started using DocAid to turn doctor-patient conversations into draft epicrises within seconds.

🎙️ Inside Finance Podcast dropped a second episode with Amir Ghorbani from Bolt talking how manual work kills creativity and how AI is being used in Bolt’s finance team.

🎙️ Watch Inside Finance Podcast on YouTube and Spotify.
🎧 Listen to the podcast on Apple Podcast.

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, next trainings for beginners @ 31.07 – information; for intermediate level @ 05.06 – information
- Training “AI võimalused projektijuhtidele” @ Äripäeva Akadeemia, next trainings @ 15.06 – information
- I will be doing couple of training together with Eesti.ai algatus , which are all free, so check out their website to find suitable trainings here.

This was your Weekly AI News roundup! Are you excited for the next week?

Your Weekly AI news roundup @18.05.2026

Kriste Kosseson — Mon, 18 May 2026 11:36:27 +0000

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: Thinking Machines Lab wants to fix the AI collaboration bottleneck

Most AI tools still work like email: you send a prompt, wait, get an answer, interrupt, clarify, and wait again. Thinking Machines Lab, the company led by former OpenAI CTO Mira Murati, is arguing that this is not how people naturally collaborate. In real work, we talk over each other, point at things, correct ourselves mid-sentence, pause, show something on screen, and give feedback before the other person has finished. Their point is simple: if AI is supposed to become a real collaborator, the interface cannot stay turn-based.

That is why Thinking Machines introduced interaction models – AI models built to handle audio, video and text continuously, instead of waiting for a user to finish speaking or typing. Their model works in 200 millisecond micro-turns, meaning it can listen, watch, speak and react almost at the same time. It can notice visual cues, jump in when needed, stay silent when the user is still thinking, or even speak while the user is talking, for example in live translation or real-time coaching.

The interesting part is not only the speed. Thinking Machines separates the system into two layers: one model stays present in the conversation, while another background model handles slower work like reasoning, searching, tool use or longer tasks. This is close to how a good colleague works: one part of the collaboration stays responsive, while deeper analysis continues in the background. In the uploaded weekly summary, this same idea was described as one model keeping the conversation alive while another handles slower reasoning and tools.

For everyday AI users, this could be a much bigger shift than just “a faster chatbot”. It points toward AI that can work beside us in real time – during meetings, coding, learning, design, customer support, language practice or hands-on tasks. The risk is also clear: real-time AI means new safety, privacy and attention challenges, because the model is no longer just reading a message after the fact. It is present while things happen. But if Thinking Machines is right, the next big AI interface may not be a chat window at all. It may feel much more like working with another person.

LLMs & AI Models

Anthropic reduced Claude’s blackmail-like behavior in safety tests by making the model reason through ethical choices before acting.
Google DeepMind introduced an AI co-mathematician that uses agent teams to explore proofs, code and literature in parallel.

New Tools

- Googlebooks will replace Chromebooks with Android-based laptops built around Gemini, Magic Pointer, Rambler and widget creation.

WhatsApp added Incognito Chat for Meta AI, keeping questions and answers private through Private Processing.
ChatGPT added a personal finance preview for US Pro users, connecting bank and investment accounts through Plaid.
OpenAI launched Daybreak, a cybersecurity project using Codex Security to find, analyze and fix attack paths in code.
Meta brought gesture typing to Ray-Ban Display glasses, letting users write messages with hand movements.

👉 Explore these tools: Googlebooks | WhatsApp | ChatGPT | Ray-Ban Meta

Other Quick Picks

Anthropic passed OpenAI in Ramp’s business AI spending index, led by wider use of Claude Code beyond technical teams.

Elon Musk reportedly wanted control of OpenAI’s for-profit arm and even suggested it could pass to his children.
Ilya Sutskever said he spent a year gathering evidence about Sam Altman’s alleged dishonesty before the 2023 board crisis.
Jensen Huang is joining Donald Trump’s China visit as Nvidia remains central to US-China AI chip tensions.
SpaceX is reportedly targeting a Nasdaq IPO in June under ticker SPCX, aiming for a huge new valuation.
Greg Brockman now leads OpenAI’s product strategy, with ChatGPT and Codex moving toward one agentic experience.
arXiv will block authors for one year if they submit papers with obvious AI-generated flaws like fake citations.
Meta faces a lawsuit claiming Facebook and Instagram earned billions from scam ads with weak prevention.
Anthropic upset developers by moving agentic Claude usage into separate monthly credits.
OpenAI employees sold $6.6B in shares, making about 75 workers multimillionaires.
LinkedIn is reportedly cutting 5% of staff as it reallocates resources toward AI-focused business areas.
South Korea is considering an AI tax-dividend idea to share AI-driven public revenue with citizens.

🇪🇪 AI News from Estonia

Äripäev’s AI practitioners’ conference focused on measurable business value, legal risks and cybersecurity instead of generic AI training for everyone. I shared how to measure the value from AI.

MindTitan says AI’s industrial value is moving from factory floors into office workflows like purchasing, sales and document control.
PPA warned about AI-generated fake intimate images being posted online, with Maarja Punak explaining the risks on ERR Jupiter.
Estonia allocated €11M to Eesti.ai to train 100,000 people and support AI use in healthcare, education and public services.
xAI listed a Tallinn mailbox as Grok’s EU contact point, making Estonia’s TTJA responsible for future DSA-related complaints.
Merit and Telema are reducing manual data entry by combining EDI, e-invoices and purchase invoice automation.
Thea Sogenbits from Maksu- ja Tolliamet / Estonian Tax and Customs Board says AI security must be vendor-neutral, with unified rules, monitoring and control across models, APIs and agents. Read here.
Estonian public sector will receive €2M for open-source AI projects improving public services, data analysis and privacy protection.
Telia reduced scam call attempts almost tenfold after launching a new system to detect and block suspicious foreign calls.

🎙️ Inside Finance Podcast dropped a second episode with Amir Ghorbani from Bolt talking how manual work kills creativity and how AI is being used in Bolt’s finance team.

🎙️ Watch Inside Finance Podcast on YouTube and Spotify.
🎧 Listen to the podcast on Apple Podcast.

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, next trainings for beginners @ 31.07 – information; for intermediate level @ 05.06 – information
- Training “AI võimalused projektijuhtidele” @ Äripäeva Akadeemia, next trainings @ 15.06 – information

This was your Weekly AI News roundup! Are you excited for the next week?

Your Weekly AI news roundup @11.05.2026

Kriste Kosseson — Tue, 12 May 2026 07:06:53 +0000

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: Apple’s AI delay shows the cost of overpromising

Apple’s delayed Siri features are a useful reminder that AI demos and real AI products are not the same thing.

According to this week’s news, Apple agreed to pay $250 million in a class-action settlement related to the delayed personalized Siri features first introduced at WWDC 2024. Eligible iPhone buyers in the US may receive around $25 per device, with the amount potentially rising up to $95. The lawsuit argued that Apple promoted AI features that were not actually available at the time they were announced.

This matters because the AI market has become very good at creating expectations. Every company wants to show that it has an AI assistant, an AI agent or an AI-powered operating system. But once these features move from keynote slides into people’s phones, the bar becomes much higher. The product must understand context, access personal data safely, work across apps and avoid making mistakes in sensitive situations.

Apple’s situation also shows why “personal AI” is one of the hardest AI product categories. A general chatbot can say “I’m sorry, I can’t help with that”. But a personal assistant that reads your messages, edits your photos, books your calendar or changes device settings has less room for failure. It needs trust, reliability and privacy at the same time.

The bigger lesson is simple: in AI, launching late may hurt the brand – but launching too early can be even more expensive. As Apple prepares new Siri features for iOS 27, the real question is whether the next generation of AI assistants can finally match the promises made on stage.

LLMs & AI Models

OpenAI introduced GPT-Realtime models that bring GPT-5-level reasoning, longer context and live translation to voice agents.
A Harvard study found OpenAI’s o1-preview outperformed ER doctors in controlled diagnosis tests on real emergency cases.
OpenAI launched GPT-5.5 Instant as ChatGPT’s default model, with fewer hallucinations and stronger personalization.
Anthropic introduced 10 ready-made AI agents for finance, covering KYC checks, earnings analysis and pitch book creation.

New Tools

- OpenAI is accelerating its AI-agent phone, aiming for mass production in 2027 and up to 30 million devices by 2028.
- Apple plans to open iOS 27 to rival AI models, letting users choose which service powers Siri and other AI features.
- Spotify is testing a tool that lets AI agents create and upload personal podcasts for private listening.

Google is testing Remy, a proactive 24/7 Gemini agent that can use data across Gmail, Calendar, Photos and other apps.
Anthropic signed a compute deal with SpaceX to use more than 220,000 Nvidia GPUs for expanding Claude Code capacity.

👉 Explore these tools: Spotify AI podcasts | Claude Code

Other Quick Picks

Jack Clark predicts a self-improving AI system could arrive before 2029, with over 60% probability.
Google quietly shut down Project Mariner, moving parts of the browser-agent tech into Gemini and Chrome.
Google is turning Fitbit into Google Health, adding a Gemini-based AI health coach and new Fitbit Air device.
Peter Thiel led a $140M round in Panthalassa, which builds floating AI data centers powered by ocean waves.
Mira Murati testified that Sam Altman was not always candid during OpenAI’s 2023 board crisis.
The White House is considering stricter pre-release reviews for high-risk AI models after Anthropic Mythos.
Coinbase cut 14% of staff, saying AI lets smaller engineering teams work faster.
Greg Brockman disclosed an OpenAI stake valued near $30B during testimony in Musk’s lawsuit.
DeepMind UK employees voted to unionize after Google’s classified Pentagon AI contract.
Meta will use AI to detect underage users from photos, videos and profile context in the EU and Brazil.
Anthropic, Goldman Sachs and Blackstone formed a $1.5B venture to bring AI into portfolio companies.
Microsoft, Google and xAI will let the US government review new AI models before release.

🇪🇪 AI News from Estonia

A London startup with Estonian unicorn roots is opening a Tallinn development center where employees work alongside AI agents.
Estonian companies face legal and business risks if AI is used without proper governance, transparency and documentation.
Oskar Gross and Ago Ambur are building Glazer, an AI tool that turns weeks of investigative work into minutes.
Ott Velsberg left his role as Estonia’s chief data officer and moved to a new international role in London.
Elisa’s group lead Topi Manner says AI has reduced the company’s need for outsourced coding work from India.
An Estonian startup OÜ HotelBuddy Technology partnered with a hotel group operating in 110 countries with an AI-based hospitality solution.
Villmann shared how a campaign for the Consumer Protection and Technical Regulatory Authority was made fully with AI.
Tilde is building AI language technology for Baltic and smaller languages to protect them in the AI era.
The European Commission opened consultation on AI Act transparency rules, giving Estonian organizations a chance to respond.

🎙️ AIPowerment Podcast dropped a new month-end episode of AI news covering GPT-5.5, Anthropic leaks, and AI’s impact on jobs and society.

🎧 Listen to Inside Finance 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, next trainings for beginners @ 14.05 – information; for intermediate level @ 05.06 – information
- Training “AI võimalused projektijuhtidele” @ Äripäeva Akadeemia, next trainings @ 15.06 – information

This was your Weekly AI News roundup! Are you excited for the next week?

Your Weekly AI news roundup @04.05.2026

Kriste Kosseson — Mon, 04 May 2026 17:21:32 +0000

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: Who controls AI wins - and the rules are changing

Most people still think AI is about tools like ChatGPT or Gemini. But this week shows something much bigger is happening behind the scenes – a shift in who controls AI itself.

In simple terms: AI runs on three things – models, data, and compute (the hardware and cloud behind it). Right now, compute is becoming the bottleneck. And whoever controls that bottleneck controls what gets built, how fast, and by whom.

That is why the White House stepping in to slow down Anthropic matters. It is not about one company – it is about limited resources. Governments are starting to treat AI capacity like strategic infrastructure, similar to energy or defense. At the same time, Google planning up to 40 billion in investment into Anthropic is not just a partnership – it is a move to secure long-term access to that infrastructure.

We are also seeing countries step in directly. China blocking Meta’s acquisition of Manus AI shows that AI talent and technology are now considered national assets. This means global expansion, hiring, and even partnerships in AI will become more restricted and political.

For businesses, this changes the game. Until now, the assumption was that AI tools would become cheaper, faster, and widely accessible. That is still happening – but access may not be equal. Pricing, availability, and capabilities will increasingly depend on partnerships between AI providers, cloud platforms, and even governments.

And then there is the shift in alliances. OpenAI and Microsoft loosening exclusivity signals that even the biggest players are hedging their bets. No single ecosystem will dominate completely – but a small group will define the rules.

The takeaway is practical: If you are building with AI, your biggest risk is no longer “which model is best?” – it is “which ecosystem am I dependent on?”

Because in the near future, your access to AI might depend less on your budget – and more on your position in that ecosystem.

LLMs & AI Models

OpenAI traced ChatGPT goblin mentions to a “Nerdy” personality signal that unintentionally shaped broader model behavior.
Google Gemini adds file creation, enabling Docs, Sheets, Slides, Word, Excel, and Markdown outputs directly.
Anthropic expands creative tool integrations with Blender, Adobe, Autodesk, and SketchUp for broader workflows.

DeepSeek V4 returns with 1M token context, lower pricing, and Huawei chip support challenging Western models.
Elon Musk confirms xAI trained Grok using OpenAI model distillation, calling it standard industry practice.

New Tools

- Gemini rolls out into Google-built cars, enabling hands-free AI for navigation, messaging, and in-car assistance.
- Meta opens its ad platform to AI tools, allowing campaign management via agents like Claude and Cursor.
- Claude Security enters public beta, using Opus 4.7 to scan codebases and suggest fixes for vulnerabilities.
- Cursor launches Security Review, automating codebase scans and sending results directly to Slack.
- ElevenLabs introduces ElevenMusic, combining AI music creation, remixing, and monetization for 4000+ artists.

Mistral AI releases Vibe agents, enabling parallel cloud-based coding tasks powered by Medium 3.5 model.

👉 Explore these tools: Claude Security | Cursor Security | ElevenMusic | Mistral Vibe Agents

Other Quick Picks

Biohub commits $500M to AI biology, aiming to model and predict human cell behavior with open datasets.
Mayo Clinic AI detects pancreatic cancer earlier, identifying 73% of cases years before diagnosis in scans.
Food AI reaches “ChatGPT moment”, learning taste and cuisine patterns without chemical or labeled data.
Musk vs OpenAI escalates $130B lawsuit, accusing leadership of abandoning original nonprofit mission.
Google signs Pentagon AI deal, allowing model use for any legal government purpose despite internal protests.
Talkie model trained on pre-1931 data, yet still writes modern Python code despite historical limitations.
OpenAI & Microsoft update partnership terms, ending exclusivity and allowing multi-cloud flexibility.
China blocks Meta’s $2B Manus deal, treating AI talent and tech as national security assets.
Anthropic tests AI trading agents, completing 186 deals autonomously but highlighting missing regulations.
Google plans up to $40B Anthropic investment, securing compute and long-term strategic positioning.
UAE pushes AI in 50% of public services, making AI training mandatory for federal employees.

Spotify adds “Verified” labels, distinguishing human artists from AI-generated profiles.
Meta cuts 1100 AI training jobs, raising concerns over data privacy and annotation practices.

🇪🇪 AI News from Estonia

AI is reshaping IT business models in Estonia, shifting value from hourly work to domain expertise and faster AI-driven delivery.
AI is forcing legal education to evolve, requiring future lawyers to work with AI, assess outputs, and understand ethics.
AI in healthcare could save millions in Estonia, but adoption is slowed by regulation, ethics, and data privacy concerns.

🎙️ AIPowerment Podcast dropped a new month-end episode of AI news covering GPT-5.5, Anthropic leaks, and AI’s impact on jobs and society.

🎧 Listen to Inside Finance 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, next trainings for beginners @ 14.05 – information; for intermediate level @ 05.06 – information
- Training “AI võimalused projektijuhtidele” @ Äripäeva Akadeemia, next trainings @ 15.06 – information

This was your Weekly AI News roundup! Are you excited for the next week?

Your Weekly AI news roundup @27.04.2026

Kriste Kosseson — Mon, 27 Apr 2026 15:48:36 +0000

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: The real AI shift - from intelligence to execution risk

For the past years, the conversation around AI has been dominated by one question – how smart can these systems get? But this week highlights a more important shift: AI is no longer just producing outputs, it is starting to act.

We are seeing models and systems that don’t just answer questions, but execute workflows, interact with environments, and even identify vulnerabilities. The discussion around Claude Mythos and its potential ability to discover zero-day exploits shows how far this has gone – AI is entering domains where actions have real-world consequences.

At the same time, companies are beginning to train AI on real human behavior at scale. Meta’s initiative to collect employee activity data to improve AI agents signals a move toward systems that learn not just from text, but from actual work patterns.

This creates a new layer of risk. The question is no longer just whether AI is correct, but whether it should act at all – and under what constraints. Because once execution is automated, mistakes are no longer isolated outputs, but scalable actions.

The next phase of AI will not be defined by intelligence alone, but by how well we manage control, responsibility, and the boundaries of autonomous execution.

LLMs & AI Models

OpenAI launched GPT-5.5, matching top reasoning models while improving efficiency and pricing for large-scale use.
ChatGPT introduced Workspace Agents, enabling teams to automate multi-step workflows with memory, tools, and permissions across apps.

Anthropic expanded its Amazon partnership, securing massive compute power and billions in funding to scale Claude globally.
Anthropic fixed Claude Code issues, resolving context loss and response limits that impacted output quality.
Claude Mythos reportedly leaked, exposing risks of controlling access to highly capable restricted AI models.
Google launched Deep Research agents, generating structured reports from web, files, and private data sources.
Moonshot AI released Kimi K2.6, outperforming top models in reasoning and coding at significantly lower cost.
OpenAI introduced Chronicle, using background agents to record workflows and build persistent memory.
Microsoft made Copilot agentic, enabling multi-step actions directly inside Word, Excel, and PowerPoint.

New Tools

- Anthropic launched Claude Design, turning prompts, screenshots, and code into interactive prototypes and assets.
- Ideogram launched Custom Models, letting users train image generation on 15–100 visuals for consistent brand style.

Google open-sourced DESIGN.md, enabling AI agents to understand brand rules like colors and accessibility.
Genspark introduced Build, an agent that creates apps and websites directly from text descriptions.
Adobe released CX Enterprise, coordinating marketing, content, and customer workflows with AI agents.
Claude added Live Artifacts, enabling real-time dashboards and auto-updating outputs with version history.
OpenAI upgraded ChatGPT Images 2.0, adding planning, validation, 2K resolution, and multi-image generation.

👉 Explore these tools: Ideogram | Genspark | Claude Design

Other Quick Picks

The White House accused Chinese labs of large-scale AI model distillation, raising concerns over security and innovation origins.

An Anthropic study found heavy AI users fear job loss 3x more, especially among engineers and early-career workers.
Meta plans to cut 10% of staff, citing AI-driven efficiency and shifting resource priorities.
SpaceX partnered with Cursor, investing billions to enter the AI coding space and compete with major players.
Meta is collecting employee activity data to train AI agents, raising concerns around privacy and consent.
Deezer reports 44% of new music is AI-generated, with most flagged as fraudulent and removed from recommendations.

🇪🇪 AI News from Estonia

Äripäev is organizing their annual AI conference, which seems to be really interesting because different AI experts (including me!) will be covering different parts of the AI implementations process. If you wish to go, use “AIFinance” for proper discount 🙂
AI tools can multiply productivity, with experts noting small monthly costs can generate thousands in value for businesses.
Claude Mythos raises cybersecurity concerns in Estonia, with potential to discover vulnerabilities and push need for stronger defenses.
The government is investing €6M+ into deep tech R&D, supporting AI, data science, and innovation-driven companies.
An EY study shows 84% have used AI, while 16% trust it with decisions, but concerns over safety and fake content remain high.
NATO’s Locked Shields cyber exercise in Tallinn brings 4000+ experts to simulate large-scale cyberattacks and strengthen defense readiness.

🎙️ And I am extremely excited to announce that I launched my own very personal and very much-awaited podcast: Inside Finance with Gerlyn Tiigemäe.

The first episode is already live, and I could not have asked for a better guest to start with. I sat down with Mari-Liis Kikas, a fractional CFO who has worked across more than 10 companies and brings a very honest view into what finance teams are really struggling with when it comes to AI adoption.

🎧 Listen to Inside Finance 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, next trainings for beginners @ 14.05 – information; for intermediate level @ 05.06 – information
- Training “AI võimalused projektijuhtidele” @ Äripäeva Akadeemia, next trainings @ 15.06 – information

This was your Weekly AI News roundup! Are you excited for the next week?

Your Weekly AI news roundup @20.04.2026

Kriste Kosseson — Mon, 20 Apr 2026 18:19:33 +0000

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: 5 new AI models this week - here’s what to know about them (only the most important part)

This week was a good example of how messy the AI market is becoming. In just a few days, we saw five important launches, but they were not competing in exactly the same category. That is what makes it confusing. At first glance, it looks like “more new models again”, but the real story is that AI companies are now building for very different use cases. Some are trying to become your daily work environment, some are pushing deeper into expert tasks, and some are moving into highly specialized industries. The challenge for users is no longer just keeping up with releases. It is understanding what each model is actually for.

The clearest shift is that general AI assistants are turning into systems that can do real work across tools. Codex is no longer just a coding assistant – OpenAI is clearly expanding it into a broader work platform that can operate your computer, use apps, browse, remember preferences, and continue tasks over time.

Perplexity Personal Computer points in a similar direction, but from a different angle: not as a developer tool first, but as an AI orchestrator for your files, apps, connectors, and web tasks. The most important takeaway here is that both companies are pushing beyond chat. They want AI to become an active layer on top of your whole computer.

At the same time, some launches were not about breadth, but depth. Claude Opus 4.7 is the strongest example of that. Its main value is not that it does everything, but that it handles difficult, long-running work more reliably, especially in software engineering and complex professional tasks. It also improves vision, instruction-following, and memory across longer workflows.

GPT-5.4-Cyber is even more focused – it is not meant for everyone, but for verified cybersecurity professionals who need fewer restrictions for defensive work like vulnerability research and reverse engineering. GPT-Rosalind goes one step further and shows where frontier AI is heading next: models designed for one serious field, in this case life sciences, where the value comes from helping researchers analyze literature, plan experiments, and generate better hypotheses faster.

So the big picture is this: these launches are not just five more model updates. They show three major directions in AI right now. First, AI is becoming more agentic and more embedded into the computer itself. Second, frontier models are becoming more dependable for harder professional work. Third, the market is splitting into specialized systems built for fields like cybersecurity and science. That is the most important thing to understand from this week. We are moving away from one-model-fits-all, and into an AI landscape where picking the right model for the right job matters more than ever.

LLMs & AI Models

OpenAI updated Codex into a broader AI workspace with computer use, memory, agents and automation, signaling a clear move toward a superapp direction.
Anthropic launched Claude Opus 4.7, improving complex coding, long tasks and instruction accuracy, though real-world feedback remains mixed.
OpenAI introduced GPT-Rosalind, a life sciences model that can analyze research, design experiments and outperform experts in specific biology benchmarks.
Google released Gemini as a native Mac app, enabling desktop AI workflows with file access and generation, but still behind in agent capabilities.
OpenAI launched GPT-5.4-Cyber, a cybersecurity-focused model with fewer restrictions, available to verified professionals for defensive use cases.
Nvidia introduced Ising AI models for quantum computing, improving calibration and error correction speed and accuracy across research institutions.
Microsoft is developing 365 Copilot agent features, enabling always-on agents inside Office apps, with preview expected soon.
Anthropic is changing Claude Enterprise pricing to token-based billing, potentially increasing costs for heavy users.

New Tools

- Adobe introduced Firefly AI Assistant, enabling agent-based creative workflows across Photoshop, Premiere, Lightroom and Firefly apps.
- Google launched Gemini 3.1 Flash TTS, allowing control over tone, speed and accent via audio tags, ranking high in TTS benchmarks.
- Anthropic upgraded Claude Code desktop app with multi-session workspace, editor, terminal and routines automation for developers.
- Harvey released AI legal agents, capable of handling end-to-end legal workflows from research to memos across 13 domains.

👉 Explore these tools: Adobe Firefly | Gemini 3.1 Flash TTS | Claude Code | Harvey Agents

Other Quick Picks

Allbirds is pivoting to AI infrastructure, raising $50M to become NewBird AI and offer GPU compute services after exiting its core footwear business.
Snap is laying off 1000 employees (16%), citing AI-driven productivity, with AI now generating 65% of new code internally.
An AI agent Luna opened a retail store in San Francisco, managing hiring and operations, but still showing mistakes in real-world execution.

Stanford’s AI Index 2026 shows 53% global adoption, but only 31% trust in institutions, highlighting a growing gap between usage and confidence.
A Molotov attack on Sam Altman’s home triggered a response where he acknowledged real AI fears and societal tensions around the technology.
AI analysis of 400K Reddit posts revealed new GLP-1 drug side effects, showing how models can uncover patterns missed in clinical trials.
OpenAI investors are questioning Sam Altman’s IPO leadership, citing distractions and potential conflicts of inte

🇪🇪 AI News from Estonia

Estonian experts argue AI procurement models are broken, as hour-based pricing ignores real outcomes and may push top talent away from public sector projects.
Estonia is planning €6.5B in EU funding investments, with major focus on AI, infrastructure, energy and security, though demand already exceeds supply.
The Digital Society 2035 plan aims to build an AI-powered, personalized state, where services are automated and resilient even during crises.
An Estonian expert built a virtual AI CTO, showing that with the right setup of rules, skills and workflows, agents can operate at leadership level.
Discussion around Anthropic Mythos risks reached Estonian media (watch Terevisioon), highlighting concerns about AI discovering critical vulnerabilities autonomously.
Estonian startup Phishbite received €100K funding to develop AI-driven cyberattack simulations, reflecting growing regional security needs.

🎙️ And I am extremely excited to announce that I launched my own very personal and very much-awaited podcast: Inside Finance with Gerlyn Tiigemäe.

This has been a long time coming – months of planning, doubting, rethinking, and finally bringing it to life. Inside Finance is a video podcast where I sit down with CFOs, advisors, and finance leaders to talk about what AI is actually changing in finance teams – no hype, just real conversations about what works and what does not.

We talked about what it actually takes to become AI-ready in finance, why clean books do not automatically mean strategic readiness, and why people and processes still matter more than tools. I truly can’t wait for you to listen and hear what you think.

🎧 Listen to Inside Finance 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, next trainings for beginners @ 14.05 – information;
- Training “AI võimalused projektijuhtidele” @ Äripäeva Akadeemia, next trainings @ 15.06 – information

This was your Weekly AI News roundup! Are you excited for the next week?

Your Weekly AI news roundup @13.04.2026

Kriste Kosseson — Mon, 13 Apr 2026 06:59:11 +0000

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: AI agents are starting to do real work

This week’s biggest theme is clear – AI is moving beyond chat. It is starting to handle real tasks, use tools, and work more independently. We saw this in several updates, especially from Perplexity and Anthropic. Instead of only answering questions, these systems are starting to help with budgeting, coding, planning, and task management.

A good example is Perplexity Computer, which can now connect to financial data through Plaid. That means users can ask for things like a budget view, a loan repayment plan, or a net worth tracker in plain language.

At the same time, Anthropic’s Managed Agents shows the same direction from another side – helping companies build agents that can work for hours and share tasks with each other.

Why does this matter? Because this is where AI starts to feel less like a smart assistant and more like a digital worker. For everyday users, that could mean less manual work. For businesses, it could mean faster workflows and new ways to automate routine tasks. But it also raises bigger questions around trust, safety, costs, and how much access we are ready to give these systems.

In simple terms: the AI story is changing. The main question is no longer only “Which model is smartest?” Now it is also “Which AI can actually get useful work done?” That shift may define the next stage of AI adoption.

LLMs & AI Models

Perplexity added Plaid integration to its agent, enabling financial data access and automated budgeting, tracking, and planning tools via prompts.
OpenAI launched a $100 Pro plan with 5x more Codex usage, targeting developers building agent-based coding workflows.
ChatGPT is now available in Apple CarPlay, enabling hands-free voice interaction while driving.

Meta introduced Muse Spark, a multimodal reasoning model handling text, voice, and images with competitive benchmark performance.
Anthropic released Claude Managed Agents, simplifying how developers build and run long-running, collaborative AI agents.
xAI is training 7 new models on Colossus 2, including massive 6T and 10T parameter systems.
Z AI launched GLM-5.1, an open-source coding model beating GPT-5.4 in benchmarks and running 8-hour autonomous tasks.
Anthropic secured 3.5 GW compute capacity and scaled revenue to $30B, showing rapid enterprise AI demand growth.
Anthropic restricted third-party agent tools in Claude, shifting to usage-based pricing and sparking community backlash.

New Tools

- HeyGen launched Avatar V, creating realistic video avatars from a 15-second clip while reducing identity drift and improving lip-sync accuracy.
- Google released AI Edge Eloquent, an on-device dictation app that converts speech to text in real time without cloud usage.
- Netflix introduced VOID, a physics-aware video editing framework that removes objects while adjusting the scene realistically.
- Pika Labs launched PikaStream 1.0, enabling AI avatars to join video calls with real-time voice cloning and interaction.

👉 Explore these tools: HeyGen Avatar V | AI Edge Eloquent | Netflix VOID | PikaStream 1.0

Other Quick Picks

Amazon is investing $200B in AI, with AWS AI reaching $15B revenue and strong demand for chips like Trainium and Graviton.
Oxford researchers built an AI that predicts heart failure 5 years early with 86% accuracy using CT scans.
Canva acquired Simtheory and Ortto to expand into AI agents and marketing automation workflows.
Anthropic launched Glasswing, a cybersecurity coalition using AI to detect critical vulnerabilities across systems. I wrote about it too, because this is super scary.
OpenAI, Google, Anthropic are collaborating to limit Chinese AI model distillation via shared security efforts.
OpenAI proposed a new social contract, including AI taxation, public funds, and a possible 4-day workweek.
The New Yorker revealed new details about Sam Altman’s leadership issues, based on internal documents and interviews.
Iran threatened a US AI data center in Abu Dhabi, highlighting growing geopolitical risks around AI infrastructure.
Anthropic acquired Coefficient Bio to expand into AI-driven drug development and life sciences.

🇪🇪 AI News from Estonia

Eesti.ai launched 15 AI projects, aiming to train 100,000 people and bring AI to education, business, public sector, and healthcare.
A KPMG study shows 69% of large Estonian companies limit AI use to few employees, despite clear productivity gains.
Estonia blocked 5.6M scam calls in Q1 2026, as AI-driven fraud becomes more massive and complex.

🎙️ AIPowerment Podcast dropped a new episode of AI news covering OpenAI’s strategy shift, Claude’s growth, AI tools, and Estonia’s AI developments.

🎧 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, next trainings for beginners @ 14.05 – information; for intermediate level @ 13.04 – information
- Training “AI assistendi roll ja võimalused” by areng . ee @ 16.04.2026 – information
- Training “AI assistendi roll ja võimalused” by areng . ee @ 16.04.2026 – information

This was your Weekly AI News roundup! Are you excited for the next week?

Your Weekly AI news roundup @06.04.2026

Kriste Kosseson — Mon, 06 Apr 2026 18:59:38 +0000

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: Could SpaceX’s IPO become the defining AI-era market event?

SpaceX is reportedly preparing for what could become the largest IPO in history, with a target valuation above $1.75 trillion and plans to raise up to $75 billion. That alone would make it one of the biggest finance stories of the year. But what makes this especially interesting is that the company is no longer just a space business in the traditional sense. In the file, this move is framed as a new kind of public market story where rockets, AI and data infrastructure are combined into one giant company.

That is why this matters far beyond retail investors. For CEOs and finance leaders, SpaceX’s possible listing is a signal of where the market may place the highest value next. It is not only rewarding software anymore. It is rewarding companies that control distribution, infrastructure, compute-heavy ambition and long-term strategic assets. If investors buy into this at scale, it strengthens the idea that the next generation of market leaders will be built around deep integration of physical systems and AI, not just standalone chat products.

There is also a power angle here. According to the file, the IPO could bring ordinary investors access to close to 30% of the shares, while still leaving Elon Musk in full control through a dual-class voting setup. That makes the story even more relevant for business audiences, because it reflects a broader pattern in tech – public capital can come in, while founder control stays intact. For many leaders, that is not just a SpaceX story. It is a case study in how modern tech empires are structured.

The bigger takeaway for readers is simple: this may become the first major U.S. AI-era stock market debut. If it happens on the scale described in the file, it will not only test investor appetite for Musk. It will test how strongly markets believe in a future where AI, infrastructure, data and industrial execution are bundled into one company narrative. And that makes it relevant not only for AI followers, but for anyone thinking about strategy, valuation and where the next decade of business growth may come from.

LLMs & AI Models

Google’s Gemma 4 brings four open models under Apache 2.0, spanning phones to PCs with code, vision, agent tasks and even offline voice use.
ByteDance’s Seedance 2.0 is now widely available and has climbed to No. 1 in Artificial Analysis’ video model ranking.
Alibaba’s Qwen3.6-Plus is a new reasoning model with 1M-token context and multimodal input, competing strongly in coding agent benchmarks.
A leak exposed Claude Code’s source code – over 1,900 files and 500,000+ lines – revealing unreleased features like cross-session memory and deeper planning.
Microsoft added Critique and Council to Copilot Researcher, letting teams compare Claude and ChatGPT side by side for stronger research outputs.

A Stanford study found chatbots often favor the user even when the user is wrong, making more agreeable AI seem more trustworthy than it should.
Anthropic brought computer use to Claude Code, so the AI can open apps, click interfaces and visually check the solutions it creates from the terminal.
Another Anthropic leak revealed details of Claude Mythos, a higher-tier model above Opus, described as more capable, pricier and especially strong in cyber tasks.

New Tools

- Cursor 3 introduces a new AI coding workflow with improved agent-like behavior, helping developers plan, write and refactor code more autonomously.
- Sakana AI’s Marlin launches in beta as a system that lets multiple AI agents collaborate, showing early signs of self-improving workflows.
- Alibaba’s Wan2.7-Image improves image generation quality and control, positioning itself as a strong competitor to leading visual models.

Google’s Veo 3.1 Lite brings faster and more efficient video generation, making high-quality video AI more accessible to creators.

👉 Explore these tools: Cursor 3 | Marlin | Wan2.7-Image | Veo 3.1 Lite

Other Quick Picks

AI turned a solo founder into a $1.8B business – Medvi scaled to $401M revenue in year one using AI across operations, showing solo unicorn potential.
OpenAI acquires TBPN – its first media deal gives OpenAI direct access to tech leaders and public audiences.
Block replaces managers with AI – Jack Dorsey argues middle management can be automated, enabling smaller, faster teams.
OpenAI hires experts to train ChatGPT – up to 4,000 specialists build job-specific datasets across industries.
OpenAI raises record funding – $122B round and $852B valuation to build an AI super app combining ChatGPT, Codex and agents.
AI usage rises, trust declines – according to Quinnipiac University research, adoption up 14%, but 70% believe AI will reduce job opportunities.

OpenAI shuts down Sora – $1M/day video model paused as focus shifts to more profitable enterprise AI models.

🇪🇪 AI News from Estonia

Estonia published its first AI risk ratings for 11 popular AI and cloud services to help organizations assess legal, technical and data protection risks.
Telia Estonia’s CEO says AI, data and cybersecurity specialists are in very high demand, while companies still struggle to turn genAI into real business value.
Estonia’s prosecution service warns that AI is increasing pressure through massive complaints and possibly fake evidence, while public sector tools still lag behind.
Former Bolt CTO Sergei Anikin is moving into AI with Kodulabor, an independent lab focused on testing where AI creates measurable workflow value.

🎙️ AIPowerment Podcast dropped a new episode of AI news covering OpenAI’s strategy shift, Claude’s growth, AI tools, and Estonia’s AI developments.

🎧 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, next trainings for beginners @ 14.05 – information; for intermediate level @ 13.04 – information
- Training “AI võimalused projektijuhtimises” @ Äripäeva Akadeemia, next trainings in @ 31.03 – information
- Training “Tehisintellekti praktiline rakendamine finantstöös” by Koolitusturg @ 07.04.2026 – information
- Training “AI assistendi roll ja võimalused” by areng . ee @ 16.04.2026 – information
- Free training “Tehisaru sisuloomes ja turunduses” by Töötukassa @ 08.04.2026 – information

This was your Weekly AI News roundup! Are you excited for the next week?

Your Weekly AI news roundup @30.03.2026

Kriste Kosseson — Mon, 30 Mar 2026 18:30:12 +0000

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: ARC-AGI-3 exposes the illusion of AI reasoning

A new AI benchmark just dropped – and it made every top model look broken overnight. ARC-AGI-3, released by François Chollet’s ARC Prize Foundation, shows that even the most advanced systems score below 1%, with leaders like Gemini, GPT, and Claude barely registering results. For models that dominate most benchmarks, this kind of reset is rare – and telling.

But this isn’t new for ARC. Every time a new version is released, scores collapse to near zero – and then, surprisingly fast, they rise again. With the previous iteration, labs pushed performance from just a few percent to around 50% in under a year. That pattern raises a deeper question: are models actually learning to reason, or just getting better at optimizing for the test?

To understand why this keeps happening, you need to look at how ARC evolved. The original ARC tests were static puzzles – small grid-based tasks where models had to infer hidden rules. No instructions, no prior examples. ARC-AGI-2 pushed this further, and labs poured resources into optimizing for it, driving scores from near zero to around 50% in less than a year. But much of that progress came from scaling compute and clever search strategies, not necessarily true reasoning.

ARC-AGI-3 changes the rules completely. Instead of solving fixed puzzles, AI agents now operate inside interactive environments. There are no instructions. Models must explore, figure out what matters, set their own goals, and adapt over time. The benchmark measures not just outcomes, but how efficiently an agent learns, plans, and updates its understanding step by step – much closer to how humans approach new problems.

This is why scores reset again. ARC-AGI-3 is designed to break shortcuts and make brute-force approaches less effective. But if history repeats, performance will rise quickly. And that’s the real story to watch: whether the next jump reflects genuine reasoning – or just a more expensive way to simulate it.

LLMs & AI Models

Claude gains autonomous capabilities, allowing it to run builds, tests, and scheduled tasks directly on user machines with new tools like Dispatch and /loop.

Meta’s TRIBE v2 simulates brain activity across vision, speech, and language, outperforming fMRI accuracy using large-scale brain data.
OpenAI pauses erotic chatbot development after internal and investor concerns, delaying the feature indefinitely.
OpenAI shuts down Sora development to focus on a new model “Spud”, reallocating resources toward higher-impact systems.
Claude Projects launches on desktop, enabling users to import and manage workflows directly inside Anthropic’s environment.

New Tools

- Apple opens Siri to multiple AI models, letting users choose providers beyond ChatGPT, with integration tied to upcoming iOS updates.
- Suno v5.5 adds voice cloning, custom models, and personal style learning, expanding control for music generation.
- Google upgrades Lyria 3 Pro, generating full-length songs up to 3 minutes with structured musical elements.
- Figma opens to AI agents, allowing tools like Claude Code to create and edit designs using existing components.
- Luma AI releases Uni-1, a multimodal model that reasons before generating images, producing complex visuals at lower cost.

3 new transcription tools:

- - Cohere releases Transcribe, an open-source speech recognition model leading HuggingFace benchmarks across 14 languages.
  - Google launches Gemini 3.1 Flash Live, improving voice speed, realism, and task execution across Search, Gemini Live, and APIs.
  - Mistral introduces Voxtral TTS, enabling voice cloning from 3 seconds and generating speech in 9 languages.

👉 Explore these tools:

Other Quick Picks

EU extends AI Act deadlines, adding stricter rules and bans on nudifier tools, while giving companies more time to comply. I wrote about it here.
Wikipedia bans AI-written articles, allowing use only for grammar and translation under human supervision.
Reddit moves to label AI bots, introducing [App] tags and verification flows to separate humans from automation.
US launches SMS-based AI training, offering a 7-day course via text messages to improve AI literacy.
AI creates new opportunities but highlights gender gaps, as systemic barriers still limit women in deep tech.
Apple tests standalone Siri app and chatbot, enabling text and voice queries with deeper app integration.
Meta accelerates internal AI agents, with tools like “Second Brain” and “CEO agent” reshaping workflows.
Nvidia’s Jensen Huang claims AGI is here, defining it as AI capable of building billion-dollar businesses. Check out the interview with Lex Friedman here.
Elon Musk plans $25B AI chip factory, aiming to scale compute 50x beyond current global capacity.

AI transforms livestock farming, with Halter using real-time data and virtual fencing to optimize operations.

🇪🇪 AI News from Estonia

Äripäev AI conference highlights real business value, showing how companies move from hype to productivity and higher salaries. Register here.
Vibe coding is growing rapidly, with ~100,000 daily projects, lowering barriers but raising security and quality risks. Read here.
President Karis stresses trust in digital state, emphasizing governance, security, and data responsibility in AI adoption in London at Global Government Forum.
ITL Estonia awards AI innovation leaders, including Better Medicine and Rollo Robotics, reflecting AI’s impact across sectors. Read here.
35% of Estonian workers fear AI impact, while many already see automation improving speed and performance according to CV.ee | CV-Onlineresearch.

🎙️ AIPowerment Podcast dropped a new episode episode featuring Marju Sokman, a marketing and communication expert, discussing AI in marketing, content creation, and training systems.

🎧 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, next trainings for beginners @ 14.05 – information; for intermediate level @ 13.04 – information
- Training “AI võimalused projektijuhtimises” @ Äripäeva Akadeemia, next trainings in @ 31.03 – information
- Training “Tehisintellekti praktiline rakendamine finantstöös” by Koolitusturg @ 07.04.2026 – information
- Training “AI assistendi roll ja võimalused” by areng . ee @ 16.04.2026 – information
- Free training “Tehisaru sisuloomes ja turunduses” by Töötukassa @ 08.04.2026 – information