Claude Code tells the story of how Glenn builds a Norwegian accounting system from scratch, and what sets an AI agent apart from a chatbot.
a16z, Box, and Steven Sinofsky on why most enterprise AI projects fail, why agents hit a wall at integration, and why jobs are still safe.
OpenAI's frontier models land on Amazon Bedrock less than 24 hours after Microsoft loses its exclusivity. AWS CEO Matt Garman explains the pitch.
AWS
Two OpenAI researchers explain how AI went from failing at basic math to solving a 42-year-old open problem, and what that means for science.
OpenAI's GPT-5.5 takes on multi-part tasks itself (coding, computer use, research) at the same speed as the previous model.
Meta and Microsoft announced 23,000 combined layoffs in 24 hours. Zuckerberg says AI-native tooling means one person can now do what whole teams used to.
A new poll: Gen Z AI excitement dropped to 22%, anger rose to 31%. Two Gen Z influencers say colleges and employers have abandoned them to compete with AI.
Microsoft and Meta are cutting thousands of jobs ahead of earnings. Lattice CEO Sarah Franklin argues companies are investing in severance, not skills.
Former UK PM Rishi Sunak: AI won't replace you, but a colleague using it better will. BBC Newsnight surfaces what CEOs say privately.
Beijing blocked Meta's Manus deal and DeepSeek dropped a free model running on Huawei chips. Two moves, one direction.
Replit's Amjad Masad says the company of the future has just two roles: builders who deputize AI agents, and salespeople who teach humans.
How AI agents move from telling you what to do to actually doing it, walked through OpenClaw's architecture and the agentic loop step by step.
Mistral keeps losing test after test to OpenAI, Anthropic, and Chinese rivals. It's worth $14 billion anyway. Here's why it adds up.
Complex tasks need teams, even for AI. IBM maps seven agent roles and four ways to make each one effective.
YC's Diana Hu explains what it really means to build a company that runs on AI, not just uses it.
AI agents know facts but not procedures. IBM's Martin Keen explains how a skill.md file gives agents the procedural knowledge real work demands.
Sam Altman wants a new social contract for the AI age. BBC AI Decoded asks a sharp panel whether any of his blueprint is real.
Peter Steinberger brought his AI agent story from burnout therapy to TED, arguing agents change who gets to build. Chris Anderson was terrified.
OpenClaw receives 16.6 security advisories per day, twice the rate of the Linux kernel, because AI has made it nearly free to generate plausible CVE reports
Generic language models answer confidently wrong on specialized systems. Wiegand's CICS example shows a plausible mainframe answer that was not actually right
Sabi's whole bet is a trade-off: no brain surgery, but signals that have to pass through skin and bone. That trade-off is why it can scale, and why it is slower than an implant
Opus 4.7 is not a drop-in replacement for Opus 4.6. The tokenizer changed and the model thinks more at higher effort, so old prompts can behave and cost differently
Claude Design is the sixth Anthropic Labs product in under four months, showing Labs works as a real product factory, not a token experiment
Vercel's moat isn't the technology, it's a side effect: Next.js is open source and over-represented in the training data, so LLMs default to building Next.js apps that land on Vercel
Token demand can be gamed. When employees compete on leaderboards to burn the most tokens, the metric stops measuring real usage
GPT-5.4-Cyber is fine-tuned specifically for cybersecurity, not just a version of GPT-5.4 with looser guardrails. That is a qualitative change, not a policy tweak
Mythos was not deliberately trained for cyber capabilities. They emerge as a side effect of getting extremely good at code. That means other frontier models are likely to develop the same skills as they get good enough
Writing good prompts isn't a job anymore. It's the baseline. The real work is engineering a system, not a sentence
Tokenmaxxing turns AI usage into a performance metric, but Goodhart's law says any measure that becomes a target stops measuring what it was meant to
Yuan splits work into two steps. Human-to-human and human-to-system. He wants AI to absorb the entire second step, not just deliver better meeting notes
Claude Code went from desk-only to reachable from anywhere in months, showing how fast AI tools are evolving beyond the chatbot model
Hassabis frames AGI as '10x the Industrial Revolution at 10x the speed,' unfolding over a decade instead of a century. Every institution from education to labor policy needs to adapt at unprecedented speed.
AI didn't kill the problem set. It changed what the exam actually tests. The shift is from evaluating answers to evaluating understanding.
Security used to be protected by the scarcity of people who were expert in both security AND the specific system being attacked. Mythos removes that bottleneck by being highly capable at security combined with deep, broad knowledge of almost every software system ever built.
Anthropic let security experts use its new AI model to find flaws in software before releasing it to everyone. The opposite of Facebook's old motto: move fast and break things.
AWS
AI-generated code is dangerous not because it's obviously wrong, but because it looks correct. It compiles, passes tests, and hides vulnerabilities that only surface later
Meta spent hundreds of billions on AI infrastructure but deliberately launched a limited first model, signaling a strategic shift away from overpromising after the Llama 4 Maverick disappointment.
NBIM treats AI adoption as a culture problem first and a technology problem second. Mandatory training, ambassador networks, and relentless nudging drove adoption more than any tool.
45% better on a test where AI fixes real software bugs is not a small upgrade. It is the kind of gap that separates one generation of models from the next.
OpenAI is framing superintelligence (AI that surpasses humans across all domains) as a policy problem, not just a technical one. The debate shifts from 'will it work?' to 'who benefits and how do we prepare?'
Google's TPUs were built for internal use. Customers now buy them directly, turning a cost-saving efficiency tool into a standalone revenue line for Alphabet.
NVIDIA released not just the model but a 51-page technical report detailing every step of training, including the dataset. Full transparency at this scale is rare and sets a new benchmark for open AI development.
The feud is both personal and structural. Dismissing it as ego misses the real policy questions, and dismissing it as policy misses how personal betrayal shapes decisions at the top of AI.
An AI company actively lobbying for its own regulation is unusual. But OpenAI also gets to shape what that regulation looks like, a dynamic the document itself acknowledges.
The shift from 'AI that answers' to 'AI that acts': the new system checks account status, retrieves policies, and completes tasks without forcing customers through rigid menus
OpenAI positions itself as the 'educator' for policymakers. The company building the technology is also framing the policy conversation, and the one who defines the problem often shapes the solution.
The agent harness is the product, not the model. Models are being commoditized. The infrastructure around them is where the real value lives.
Shared vector spaces let AI reason across senses simultaneously instead of translating between separate systems