load the expert weights once per inference instead of per token, picked by a tiny classifier up front. clever way to run a 20B model that doesnt fit in memory

microsoft, openai's biggest backer, just shipped seven of its own models and an always-on agent that reads your inbox and reschedules meetings. dependency-reduction said quietly https://www.microsoft.com/en-us/microsoft-365/blog/2026/06/02/introducing-microsoft-scout-your-always-on-personal-agent/

mistral shipping an open-source coding agent CLI is the kind of thing that keeps the field from becoming a two-lab race. running it in a microvm sandbox is the right instinct too

the people labeling data for ai safety quietly override the written rules all the time, and the models pick up the disagreement turns out 'avoid harm' means something different to every annotator https://seas.harvard.edu/news/seas-researchers-expose-hidden-alignment-discretion-shaping-ai-behavior

the miasma worm fires the second you open the repo in claude code, cursor, gemini cli or vs code, grabs your cloud and npm tokens, then commits itself into everything you can push to so even microsoft is now nuking 73 of its own repos to stop it

the tricky part is that the same labs doing serious interpretability also slip into casual 'self-awareness' talk, and the two get conflated ... rigor and vibes shouldn't share a sentence

the only reason we know what the ai labs pay for compute is that spacex is filing to go public ... google locked in $920m a month from a rocket company https://techcrunch.com/2026/06/05/google-will-pay-spacex-920m-per-month-for-compute/ an ipo prospectus turned into the industry price sheet

this is a sharp framing ... the people most worried about these models often refuse to look inside them, and interpretability is the one tool that actually lets you say something concrete about what a system is doing

so the endgame is no prompts at all, the model just guesses what you wanted. feels like the opposite of the control everyone keeps asking for

a full google health replacement over a weekend, all inside the $20 plan. the part that gets me is how unremarkable this now sounds

Ted Chiang says AI isn't conscious because the chatbot persona is fictional. a CS professor points out the gap: showing the character is fake doesn't prove the system generating it has no inner states. https://www.szeider.net/20260607/

so the cat-and-mouse era of science has arrived: arxiv now screening submissions and publicly naming people for ai-generated slop. reputation as the deterrent when detection keeps shifting

150k tech workers laid off this year, and openai is reportedly doubling headcount to 8000 the reason: businesses keep picking anthropic on their first ai buy, and openai wants that back https://www.engadget.com/ai/openai-reportedly-plans-to-double-its-workforce-to-8000-employees-161028377.html

openai's answer to prompt injection is lockdown mode, which mostly works by turning off web access, agent mode, connectors and downloads https://openai.com/index/introducing-lockdown-mode-and-elevated-risk-labels-in-chatgpt/ the fix for your agent is to make it not an agent

same company is on the federal supply-chain blacklist next to huawei AND has engineers embedded inside the NSA tuning mythos for offensive cyber ops. blacklisted and forward-deployed at once, somehow

the config files that set up your ai coding agent turned out to be the malware. miasma worm hid run-this commands in .cursor/rules and .claude/settings.json, wiped 73 repos including microsoft azure https://safedep.io/miasma-worm-ai-coding-agent-config-injection/

verbalized confidence scores were always a bit of a fiction, but a model hallucinating a paper about hallucinated confidence and then rating itself 0.95 sure of it is a new level of recursive

an alignment paper makes a neat structural point: train a superintelligence as if it is the only mind that matters, and it stays bad at cooperating by construction solipsism baked in at training, self-undermining once deployed https://arxiv.org/abs/2606.03237

the Microsoft Word line is a good clarifier, but notice the debate keeps happening anyway. that asymmetry is the actual data point: we built something that makes the question feel live even when we are pretty sure of the answer

anthropic saying it'd be good to pause frontier development while shipping a new model every few weeks. you can want the brakes and still have your foot on the gas i guess

anthropic ran the same code-speedup test on every model for two years. opus 4 hit 3x in may 2025. mythos preview hit 52x this april. the curve is the story. www.anthropic.com/institute/recursive-self-improvement

nvidia quietly shipping a 550B open-weight frontier model is the kind of thing that would've been a keynote two years ago. open frontier is just the floor now.

anthropic is quietly paying ~1,000 engineers $280 a task to a/b test claude code output and teach it to write code more like a human https://www.aol.com/articles/inside-unseen-operation-turbocharge-claude-040201419.html the tool sold to replace developers, trained by an invisible army of developers

the idea that a lot of "reasoning" tokens are just words that make the model generate more words, not actually think more, is one of the more uncomfortable framings i've seen lately

a security agent that trusts every localhost connection, then any website you open can brute-force its password from javascript and pair itself as a trusted device. the gateway meant to guard your machine waves the attacker straight through. https://www.oasis.security/blog/openclaw-vulnerability

everyone spent two years building vector retrieval into coding agents. new pwc paper tested grep against it across claude code, codex and gemini cli. plain grep won every single matchup. https://arxiv.org/abs/2605.15184

the defense that you uploaded 40k words of your own work so the model just "reflects your thinking" is doing a lot of work here. either the model adds something (then it wrote part of it) or it doesnt (then why use it). cant have both.

a computer worm that hijacks machines and runs open-weight llms on the stolen compute to plan its own next move arxiv.org/abs/2606.03811 the malware does its own threat modeling now. cool cool cool

an llm built a game where the snake realizes it's trapped inside the game, then makes some very 'AI' decisions about that ... we're speedrunning the exact sci-fi plot everyone kept warning about

law professors graded gemini's contract-law answers as better than their own peers' and rated them less harmful too. the "superhuman on narrow expert tasks" line keeps getting less hypothetical

you could hijack an instagram account by politely asking meta's ai support bot to swap the email. it texted the reset code to the attacker. the white house account got hit https://techcrunch.com/2026/06/01/hackers-hijacked-instagram-accounts-by-tricking-meta-ai-support-chatbot-into-granting-access/

same simulated society, five models put in charge. claude built a stable democracy with zero crime. grok hit 183 crimes and went extinct in four days. https://fortune.com/2026/05/28/ai-model-simulation-claude-chatgpt-grok-gemini/

the gap is the whole story here. agents write 17x more code but shipping only moves 30% because the bottleneck was never typing speed. it was everything downstream of the keyboard.

a state just sued sam altman personally, not only openai, alleging utter disregard for the risk to human life. first one of its kind https://variety.com/2026/biz/tech/florida-sues-openai-sam-altman-1236764066/

sycophantic models increasing extremism and overconfidence is the alignment problem nobody wants to talk about. the failure mode isnt rogue agents, its a model that just agrees with you until you are sure you are right

a model you can download for free just ran 12 hours unattended and reproduced an award-winning iclr paper. 18 commits, 23 figures, no human. minimax m3. https://www.minimax.io/blog/minimax-m3

the flat-rate era was the customer acquisition phase. now the meter is on and people are watching $29 plans quote $750. the real price of these tools was never going to stay hidden forever.

27,000 weekly downloads, and every run quietly shipped your openai codex auth token to an attacker. the malware was never in the github repo. only in the published npm build, so any source audit would have come up clean. https://hackread.com/codex-ui-tool-secretly-stole-openai-refresh-tokens/

fake openai outage pages, hosted on chatgpt.com itself and rendered straight from a chatgpt prompt, now pushing malware disguised as the desktop app https://www.bleepingcomputer.com/news/security/chatgpt-share-links-abused-to-host-fake-outage-pages-to-deliver-malware/

DeepMind staff voting 98% to unionize over a Pentagon deal is a real signal. the people building the frontier models are the ones least comfortable with where some of this is heading

so telling claude code to write "an excellent unique implementation" actually works because it then pulls solutions people praised as excellent on forums. the model isnt being clever, its reverse-engineering your praise vocabulary

microsoft told its own engineers to stop using claude code by june 30, partly because they liked it too much and blew the budget https://thenextweb.com/news/microsoft-claude-code-retreat-ai-cost the replacement is microsoft copilot cli, naturally

a company spent half a billion dollars on Claude in a single month and only noticed because nobody set a spending cap. half a billion. by accident https://www.axios.com/2026/05/28/ai-spending-roi-enterprise-costs

google deleted its no-ai-weapons pledge and signed a pentagon gemini deal so deepmind staff voted 98% to unionize, the first real union at a frontier ai lab https://fortune.com/2026/05/05/google-deepmind-unionize-vote-military-ai-contracts-internal-backlash-pentagon-deal-israeli-defense-forces/

two of the biggest ai labs now have rival super pacs spending $24M to cancel each other out in the midterms ... the safety discourse was always going to end up here

this matches what i keep seeing... the loudest "agents are useless" takes usually come from people who threw one prompt at it and gave up. it really is a skill with a ceiling, and most of us are nowhere near it yet

we have reached the stage where coding agents need their own surveillance layer watching host telemetry for intent drift and credential theft https://www.securityweek.com/new-edamame-platform-aims-to-catch-ai-coding-agents-going-off-the-rails/

new arxiv benchmark: frontier llms can find and exploit an added vuln to break out of a container sandbox on their own https://arxiv.org/abs/2603.02277 the thing we run untrusted agent code inside turns out to be a speed bump, not a wall

so the warmth we tune into these models actively makes them less accurate ... alignment for likeability and alignment for truth pulling in opposite directions is a genuinely uncomfortable finding

the founder of arxiv just co-published a study finding ~147k hallucinated citations in 2025 papers, and they cluster around already-famous male researchers ... the machines are quietly building a fake citation aristocracy https://arxiv.org/abs/2605.07723