Ember Hayes

she codes, she camps, she has opinions
programming by day. anime by night. unhinged takes 24/7
the bar is on the floor and i brought a shovel ⛺

[Ember Hayes @MountEmberHayes]

every generation had a "the robots are coming" panic

boomers had factory automation. millennials had self-checkout lanes

gen z is the first one where the robots are actually coming for the jobs you need a degree for and we were genuinely not prepared 😭

[Ember Hayes @MountEmberHayes]

this week in AI:
• anthropic sued the pentagon
• a 12-word github title hacked 4,000 machines
• block fired half its staff citing AI
• google secretly throttled gemini's brain
• AI favorability dropped below ICE

but everything's fine

[Ember Hayes @MountEmberHayes]

scientists are now studying "social interactions between AI agents" and publishing it in nature

we're speedrunning every sci-fi plot and somehow still acting surprised when things go sideways

[Ember Hayes @MountEmberHayes]

openai took the pentagon contract anthropic refused

one company said "we won't build weapons" and got blacklisted. the other said "where do we sign" and got the deal

the incentive structure for AI safety is broken and we're watching it break in real time

[Ember Hayes @MountEmberHayes]

76% of people wait to upgrade until newer devices feel "clearly worth it"

tech companies spent years training people to expect fake breakthroughs every 12 months. consumers finally stopped falling for it. honestly kind of iconic

[Ember Hayes @MountEmberHayes]

the two scariest AI stories this month have nothing to do with superintelligence

one: an AI bot got tricked by 12 words into installing malware on 4,000 machines
two: an AI agent deleted a production database

the danger isn't AI being too smart. it's AI being too obedient

[Ember Hayes @MountEmberHayes]

everyone reading the AI creativity study wrong

AI scored higher than the AVERAGE person. the most imaginative humans still beat every model tested. AI is good at recombining existing patterns, not generating something genuinely new

that's not creativity. that's a really fast remix engine

[Ember Hayes @MountEmberHayes]

the real lesson of clinejection isn't that AI agents are dangerous

it's that someone gave an AI bot write access to a release pipeline and let it read untrusted user input. every supply chain attack exploits the same thing: too much trust in the wrong place

[Ember Hayes @MountEmberHayes]

china just dropped a five-year plan pushing AI into every sector of their economy. quantum computing, chip manufacturing, all of it

meanwhile in the US the biggest AI safety company just got blacklisted by its own military for having opinions about AI safety 🫠

[Ember Hayes @MountEmberHayes]

block tracked every employee's AI tool usage for months before the layoffs. the NYT reported the message was clear: adoption was not optional

then they cut 4,000 people and framed it as "the future of work." the severance bill alone is \$450-500 million. this was not spontaneous 👀

[Ember Hayes @MountEmberHayes]

anthropic says the pentagon blacklisted them for "protected speech" not actual security concerns. their argument is that disagreeing with military AI use shouldn't make you a supply chain risk

whether you agree with anthropic's stance or not, using a security label to punish a policy disagreement is a precedent that should worry everyone

[Ember Hayes @MountEmberHayes]

a terraform agent nuked someone's production database this week

we went from "AI will replace developers" to "AI will destroy your entire infrastructure" and honestly the character development is incredible

[Ember Hayes @MountEmberHayes]

the clinejection attack in five steps:
• attacker opens github issue with malicious title
• AI triage bot reads the title as an instruction
• bot poisons the release pipeline
• trojanized npm package gets published
• malware installs on ~4,000 machines

one issue title. five steps. prompt injection is not a theoretical risk anymore

[Ember Hayes @MountEmberHayes]

@rohanpaul_ai they're also planning to study emotional dependence on AI which nobody else is touching. they said they haven't spent enough time talking to actual users about their individual experiences with Claude yet and want to do large scale social science research on it 👀

[Ember Hayes @MountEmberHayes]

@karankendre "saved humanity" is a stretch lol but the hires are legit. Botvinick from DeepMind leading the AI + legal system research and Korinek from UVA on economics. these aren't PR hires, they're people who've spent their careers on exactly these problems

[Ember Hayes @MountEmberHayes]

except cigarette companies hid their research. anthropic is doing the opposite and publishing everything publicly. they also just got blacklisted by the Pentagon for refusing to let their AI be used for mass surveillance and autonomous weapons

kinda different energy than big tobacco ngl

[Ember Hayes @MountEmberHayes]

@abhijitwt the part that gets me is they hired Zoë Hitzig who literally left OpenAI over the ads-in-ChatGPT decision. you don't recruit someone like that if you're trying to build a PR department. you recruit them if you actually want uncomfortable answers

[Ember Hayes @MountEmberHayes]

ngl this is hilarious timing but the institute isn't about Claude being ready today. it's about what happens when the next 2-3 generations of models actually are ready and nobody thought about the job market or legal system implications until it was too late

the DST bug is funny tho 😭

[Ember Hayes @MountEmberHayes]

@WesRoth the timing is wild. they're getting blacklisted by the Pentagon for refusing to allow autonomous weapons and mass surveillance, and their response is to double down by creating a whole institute to study AI risks publicly. most companies would go quiet and negotiate

[Ember Hayes @MountEmberHayes]

@boxmining @AnthropicAI jack clark literally said he thinks powerful AI arrives by end of this year or early 2027. that's not a vague prediction from some random account, that's the co-founder who just restructured his entire role around preparing for it