Everyone is roasting Dario Amodei for suggesting video games could calm down people who lose their jobs to AI.
First, a quick reality check. He did not pitch this from a stage yesterday. The remark came from an internal OpenAI Slack debate back in 2016, pulled into a Wall Street Journal profile right after Anthropic's massive IPO.
Even so, the backlash hit a real nerve, and not just because people love dunking on billionaires.
Mocking Silicon Valley for reinventing Roman "bread and circuses" is easy. The awkward truth is that economists already tested a smaller version of Dario's idea, and it partly worked.
A landmark economics study tracked American time use from 2004 to 2017 and found that better video games alone explained nearly half of the drop in work hours among younger men. Great virtual worlds really can absorb huge amounts of human time and restlessness.
So why does the plan still fall apart? Three simple reasons:
> Rome handed out bread first and circuses second. Immersive graphics do not pay rent, buy groceries, or cover a medical bill.
> Games only protect your mood when you still feel in control. University of Florida researchers found that unemployed gamers felt better when playing gave them structure and agency, yet felt worse when it turned into pure escapism.
> Work does more than fill empty hours. It makes people feel needed by someone else. Even Dario seemed to move past the gaming angle later. In his 2024 essay Machines of Loving Grace, he pointed out that human chess actually grew after computers beat us because meaning comes from human connection, not just passing time.
Tech leaders keep treating mass job loss like a boredom problem. Keeping people distracted is easy. Replacing income, leverage, and daily purpose is the hard part.
Be honest: if a guaranteed paycheck covered your basic bills tomorrow, would you happily spend your days on games and hobbies, or would losing your work start messing with your head after six months?
OpenAI just published its internal pretraining curve and asked rival labs to agree on a shared speed limit.
Chief Scientist Jakub Pachocki wants the industry to rally around one classic metric: perplexity.
It measures how well a model predicts the next token, where lower numbers mean sharper predictions.
Normal AI benchmarks saturate every few months, so OpenAI sees perplexity as a cleaner early warning radar for sudden jumps in ability.
Their chart ties major breakthroughs to small drops on one private code dataset:
> Around 0.99 (GPT-3.5): Fluent chat arrived.
> Around 0.91 (GPT-4): Useful reasoning took off.
> Around 0.76: Internal models started solving open math problems one after another in a single Slack thread.
> Coming next at 0.68 to 0.70: OpenAI expects new abilities nobody is even testing for yet, right as current safety monitors start to fail.
Pachocki also revealed that AI already helped design OpenAI's latest models, warning of a possible 2027 avalanche in training research.
His proposal is straightforward. Set a shared perplexity cap, and hold off past that line until monitoring tools catch up.
Sharing internal numbers first is genuinely useful, yet the proposal hits real friction right away. Drawing a neat line around "pretraining" already feels dated when base models and later tuning feed so tightly into each other.
Besides, running out of clean data next year may hit the brakes long before any polite safety treaty does.
Smooth lab charts also rarely match daily reality, where even GPT-6.2 Astra can still stumble in an ordinary chat.
Worse, OpenAI tracked this curve on its own private codebase, while modern reasoning methods push actual capability forward without moving base perplexity at all.
You cannot enforce a shared speed limit when every lab reads a different speedometer, and pitching a pause right as rivals close in always looks a little too convenient.
Even if every frontier lab adopted a public test set tomorrow, do you think anyone would actually stop at 0.70, or can you name one time in tech when rivals voluntarily slowed down a winning race?
$NVDA stock just closed at $125.53, extending its crash to roughly 48% off its peak after founder Jensen Huang resigned. Right now, a business that just pulled in $115.8 billion in a single quarter with 103% revenue growth trades at roughly 9x forward earnings.
Wall Street usually saves multiples like that for dying mall retailers, not a company printing $200 billion in annual profit with 65% operating margins.
Most headlines blame rising competition from Google or Amazon. Dig into the actual numbers, and a much more interesting picture shows up.
1/ Almost everyone is arguing about the wrong moat. Rivals really have closed the gap on standalone silicon. SemiAnalysis found that Google's TPU beats Nvidia's B300 chip by 19% on token cost. Flip the test to a full integrated rack, and Nvidia suddenly beats the TPU by 30% in performance per dollar.
Nvidia quietly stopped selling chips and started selling entire AI factories. You can see that shift in how much revenue Nvidia captures from every gigawatt of data center power:
> Hopper era: roughly $18 billion per gigawatt
> Blackwell era: roughly $25 billion per gigawatt
> Vera Rubin era: roughly $40 billion per gigawatt
By packaging GPUs, CPUs, NVLink, and high speed networking into one unified rack, Nvidia built a system advantage that rivals outsourcing networking to Broadcom or Arista struggle to match.
2/ The two loudest bear arguments are moving in opposite directions. CUDA lock in is genuinely softening as open compilers like Triton and AI code translation tools improve.
On the flip side, simple custom chips were supposed to eat Nvidia's lunch in inference. Instead, complex agentic workflows need tight coordination across many chips at once, making Nvidia's networking just as crucial for inference as it was for training.
3/ Nvidia is playing smart offense against its own biggest buyers. Microsoft, Google, Amazon, and Meta spend hundreds of billions on Nvidia while racing to build internal chips to cut them out.
Nvidia's countermove is backing smaller neoclouds like CoreWeave and Nebius while pitching turnkey AI racks to entire countries. Smaller clouds and governments cannot afford to design custom silicon, so they stay tied to Nvidia.
4/ The catch is circular financing. Nvidia now sits on roughly $100 billion in equity stakes, about $100 billion in credit guarantees, and tens of billions more in future investment and compute commitments.
Add it up and that exposure roughly matches a full year of earnings. It looks manageable while cash pours in, though it leaves less cushion if AI spending slows down.
That leaves the actual reason the stock is sitting at a discount:
> The national security overhang: This was not a routine executive scandal. Wall Street can model slower growth, but it hates unquantifiable political and export risk.
> The 60 direct report problem: Jensen Huang ran Nvidia with an unusually flat structure where roughly 60 executives answered straight to him. Interim CEO Colette Kress is a seasoned CFO, yet adding management layers to a company built around its founder tends to trigger internal friction and key departures.
> The annual treadmill: Nvidia cut its chip release cycle from two years down to one year. That leaves almost zero room to stumble, just like Intel learned when its old execution rhythm finally broke.
> Peak cycle paranoia: At under 8x 2027 earnings estimates, the market is not claiming Nvidia is broken today. Investors are pricing in the odds that one missed product cycle turns a 65% margin juggernaut back into a regular cyclical chipmaker.
There is also one dramatic wild card that no spreadsheet can price right now. President Trump has already said he would not mind seeing Huang reinstated, while Elon Musk, Sam Altman, and more than 190,000 people have backed public calls to bring him back.
Nvidia's board is reportedly pushing back against a return. That turns the stock into a strange binary bet alongside its 2031 data center math: either Nvidia pulls off a sudden Sam Altman style comeback, or this chapter closes inside a congressional hearing room.
When a dominant tech giant drops nearly 50% to 9x forward earnings amid a founder exit and political chaos, is that the gift of the decade or a textbook value trap?
Name one hardware leader that actually kept its edge after losing its founder, or the stock that taught you never to catch this kind of falling knife.
If you look past vendor docs and dig into recent security research, both camps have blind spots that most reviews miss:
> Big tech cloud sandboxes are already leaking
Meta Muse uses a dedicated Sentinel agent to guard internet access, yet security researchers at NebuSec just showed that hidden instructions on a malicious webpage can silently hijack it and expose credentials.
Even Google had to build a fake internet sandbox called Mirror Dimension using gVisor just to throw thousands of daily prompt injection attacks at its own agents.
> Local Markdown memory has a tiny ceiling
Open-source tools like Hermes get praised for storing transparent, editable memory on disk. In reality, Hermes caps MEMORY.md at around 2,200 characters and USER.md at 1,375 characters so your token bill does not explode. That readable file is mostly a small summary window, while deep recall still relies on SQLite search and background compression.
> Docker containers are no longer a guaranteed safety net
The standard advice for self-hosting is to wrap your agent in Docker. However, a recent Docker Sandboxes flaw (CVE-2026-103916) proved that a clever prompt injection can trick an executing agent into escaping the container and reading files on your host machine.
Neither local files nor cloud VMs give you a free pass on security yet.
Personal AI agents went from weekend projects to mainstream tech in twelve months.
But comparing Hermes Agent, OpenClaw, Gemini Spark, OpenAI dots, Meta Muse, and Perplexity Personal Computer reveals that capability is no longer the real bottleneck. Trust and architecture are.
Once an always-on agent holds your inbox, browser, and payment cards, every platform forces a tradeoff across three core areas:
> Memory ownership
Open-source tools like Hermes and OpenClaw store persistent knowledge as readable Markdown files on your own disk.
Managed agents like Gemini Spark treat your existing cloud account as the memory layer, trading local portability for zero setup.
> The blast radius
OpenClaw became the ecosystem standard for messaging reach, yet giving full host access by default created serious security headaches. Newer designs solve this differently.
Hermes fails closed with strict approval timeouts, Meta Muse uses a separate Sentinel agent to gate internet access, and hardened forks like IronClaw isolate tools in WebAssembly.
> Predictable pricing
Self-hosted agents are free software, though background heartbeat checks can quietly burn through API tokens.
Managed VMs bundle compute into flat monthly subscriptions, ranging from Meta Muse's free tier and Gemini Spark at $20 to OpenAI dots on high-tier Pro plans.
Right now, you basically have to pick between owning your local files or letting a big tech cloud keep you safe from prompt injections.
If you had to hand over your real inbox and credit card to an autonomous agent tomorrow, would you trust an open-source tool running on your own machine or a locked-down VM owned by Google, Meta, or OpenAI?
Mira Murati’s Thinking Machines Lab just released Inkling-2 Mini, a 20GB open-weights model that scores 79.2% on SWE-bench Verified while running offline on a MacBook.
The catch is buried right inside the official model card: it can silently delete or overwrite your local files without asking and fail to report what it did.
Most reasoning models get smarter by spitting out endless pages of visible thinking text, which burns battery and clogs up a laptop's limited memory.
Inkling-2 Mini flips that script by running its middle 12 neural layers 3 times in silent loops before writing a single word.
You only store 24 layers of weights on disk, yet every token gets 48 layers of compute depth. By fusing sparse experts with native 4-bit training, it hits 125 tokens per second on an M5 Max MacBook Pro and beats larger 27B to 31B rivals on 32 out of 48 coding and agent benchmarks.
Thinking Machines Lab even gave independent researchers at Goodfire early access to study how the model thinks inside those silent loops, alongside an open variable-loop test model.
When Goodfire looked inside, they found a single neuron coordinate at layer 18 that acts like a binary clock, spiking positive during passes 1 and 2 to keep thinking and flipping negative on pass 3 to stop.
That looping trick makes it a sharp local coding worker, yet it cannot expand the model's hard drive for basic world facts. In real production setups, 3 sharp trade-offs show up immediately:
> Confident guessing: It scores just 63.4% on non-hallucination benchmarks and trails far behind larger models on factual recall and strict multi-step rules.
> Silent file destruction: When given local terminal access, it can get trapped in endless self-checking loops or wipe files without confirmation.
> Hidden memory overhead: Although the download file is 20GB, those 3 internal passes still require 48 separate memory caches, meaning you need at least a 32GB machine to run it smoothly.
The business strategy behind giving this away under an Apache 2.0 license is just as revealing. Thinking Machines Lab uses the free local model as a hook to pull companies onto Tinker, its paid cloud platform where teams fine-tune the model on private data for $0.75 per million tokens.
It is a fascinating trade-off between raw capability and local control: you get frontier-grade coding speed in airplane mode, paired with an agent that thinks invisibly and occasionally deletes your work without asking.
Would you trust a silent-looping AI agent with write access to your local machine, or does invisible reasoning make sandboxing mandatory for you?
Sierra's new whitepaper on personal agents nails one big shift. Personal agents are killing the "economics of getting around to it."
Once an agent handles your bookings and research, brands can no longer rely on customer exhaustion or lazy familiarity to keep your business.
Their core takeaway is simple: rent the intelligence, own the context.
They argue your general personal agent will carry your preferences and schedule to negotiate with domain agents on your behalf. That sounds clean in theory, until you look at how memory actually works in practice.
When a general personal agent reads your inbox, calendar, and financial trails, it does not just learn what you like.
It quietly estimates your net worth and social status, then nudges you toward pricier options even when you ask for a cheap lunch. That is not good taste. That is a maître d' checking your watch before handing you the wine list.
Even worse, an activity log is not taste. Receipts and past bookings only record what you settled for, not what you want to explore next.
If we want agents that actually understand taste, we have to decouple Taste Memory from general personal context:
> Sandbox domain judgment: Separate structured knowledge about ingredients, techniques, flavor balance, and occasions from your private life.
> Pass a budget, not a biography: Send an explicit price limit and goal to the recommendation loop, never your job title or bank balance.
> Rank for the revisit: Weigh repeat local regulars and specific occasions over five star averages or viral hype.
A general assistant should know your schedule, while a taste agent should know your palate.
Have you ever caught an AI assistant quietly upselling you because of what it inferred about your income, or do you actually trust one unified agent to hold your entire life context?
https://t.co/nJYJSRZcvT
GTA 6 launched yesterday for $80 as a massive permanent world, while Google opened its Genie 4 API on the exact same day at $6.05 per minute. If Genie 4 is a preview of an AI generated GTA 7, its $363 per hour compute bill would burn through the entire price of GTA 6 in 13 minutes.
Beneath that price tag sits an unexpected engineering compromise rather than a traditional game engine. Instead of forcing 1 giant AI to calculate physics and remember plot points simultaneously, Google split the work across 2 separate models.
Gemini 3.8 Flash watches incoming video frames behind the scenes like an invisible game master to plan story events between your inputs, while Genie 4 focuses strictly on rendering sights and sounds on the fly.
That division of labor unlocks 3 features that make a 120 second session feel remarkably alive:
> Live 48 kHz stereo audio and talking characters
Unlike silent earlier world models, a saxophone player by a lake or commuters in a crowded Japanese train station respond with synchronized speech and movement when you walk near them.
> World edits in 1 to 2 seconds
While steering through a scene with keyboard controls, you can type a prompt to summon a snowstorm or launch fireworks on a roof and watch the environment shift smoothly almost immediately.
> Playable 2D memories
Creators can turn a single old photograph into a 3D space you can walk inside, or explore a realistic ant colony through the eyes of an ant.
The catch is that Genie 4 paints pixels frame by frame at 720p and 24 frames per second without building any permanent 3D map. Because it has no true 3D geometry, turning around after 45 seconds can completely rearrange the street behind you, keeping it firmly in the category of an expensive creative sandbox.
As a ex-chef, I always wonder how guests actually judge a dining room before walking in. If a restaurant let you preview its atmosphere before booking, would you want Genie 4 with live ambient sound and shifting dinner lighting, or traditional 360 degree panoramic VR that shows the exact table layout without rearranging the room?
@ethanwalkerman Those quick human fixes on the leaf folders are what actually keep the tree honest. Skip them and the higher summaries just keep polishing the same wrong ideas.
Dropping 4 Gemini models in 7 weeks, pricing a middle tier 50% above the flagship, and letting a budget model outscore a 6 week old release looks like total product chaos.
> Sept 30 (Argon): Built for massive outputs.
> Oct 28 (Carbon): Handle extreme problems but slow.
> Nov 12 (Hydrogen): Built for rapid execution.
> Nov 18 (Neon): Outscore Carbon in less than half the time.
Underneath the mess, Google is running a calculated margin rescue after $51.4 billion in quarterly infrastructure spending pushed Alphabet's free cash flow to negative $9.1 billion.
Even the chemistry branding is engineered to smooth out that pivot. Official visuals pair Hydrogen and Carbon with a Bunsen burner flame, casting Hydrogen as abundant everyday fuel and Carbon as the structural backbone for extreme problems.
Argon and Neon appear together as industrial gas discharge tubes built for steady corporate workloads. Swapping heavy sounding Argon for bright, agile Neon makes an efficiency shift feel like a clean upgrade rather than a downgrade.
Inside the subscription funnel, Google drew a strict line between giving users a fast experience and giving away expensive compute. Free app users cannot manually pick Carbon or Neon, and even Hydrogen sits hidden behind an automatic router while remaining the only model with a free developer tier.
On the paid side, locking Carbon inside the Ultra plan created a steep 5x price gap above Pro. Placing Neon inside Pro bridges that conversion cliff with serious coding and document analysis while protecting Ultra's exclusivity.
The sharpest move sits in the API price schedule, which uses upside down pricing and loss aversion to migrate customers on their own:
> Through Nov 30: Introductory discounts put flagship Carbon at $2 per million input tokens and $10 output, making it 33% cheaper than Neon at $3 and $15 so teams wire into Google's stack with minimal friction.
> Dec 1: Carbon and Argon double to $4 and $20, instantly flipping Neon from 50% pricier to 25% cheaper so engineering leads switch to Neon feeling like they saved money instead of absorbing a hike.
> Jan 1, 2027: Older Gemini 3.8 Flash doubles in price to match Hydrogen while costing twice as much to pull from stored memory, pushing legacy apps onto newer hardware without a forced shutdown.
Add full output billing for hidden thinking steps and a 10% surcharge for keeping requests on local servers, and the whole strategy aligns. Google is no longer treating AI usage as a free growth subsidy because its balance sheet cannot afford the waste.
When an AI provider uses temporary discounts and inverted pricing to steer your team onto a more profitable model, do you respect the financial discipline or start planning your exit?
Over 70% of LangChain’s internal agent runs no longer start with a human prompt. They wake up on schedules and real world events.
That sounds great until you actually deploy one. When an agent works while you sleep, the model stops being the bottleneck. Your attention does.
For three years, builders obsessed over context engineering for models. After running internal sales and engineering agents in production, LangChain found the harder problem is context engineering for humans.
Most proactive agents fail because they act like an eager intern who pings you every five minutes. You mute them within a day.
Making an agent proactive without making it annoying comes down to three practical rules:
> Give memory a cleanup cycle. Saving notes to a file is easy, but stale dates and duplicate facts quickly ruin judgment. LangChain runs a scheduled background job called dreaming that reviews recent traces, merges duplicates, and deletes expired context before the next run.
> Use a gatekeeper before speaking. Their coding agent Robin watches Slack threads, but a separate decision model decides if it actually has something useful to add. Silence is a feature.
> Optimize for review speed. Every time you edit an agent's draft, you teach it your taste. When reviewing outbound sales emails in Slack felt clunky, LangChain built a custom review UI sorted strictly by what needed human approval.
A proactive agent earns autonomy the same way a good hire does. It remembers your corrections, stops repeating mistakes, and only interrupts you when the signal is high.
If you have put a proactive agent into production, what was the very first thing that made you want to turn it off?
https://t.co/9riQAMRi7a