I spent years cooking for people who said, “Anything is fine.”
It never was.
So I’m building Mate’s Table—turning your taste into an agent, because a 4.7-star rating still has no idea what you want for dinner.
https://t.co/EswJSO4xN4
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?
We decided to provide more transparency into pretraining progress at OpenAI for the sake of moving toward shared safety bars. In “Perplexity and Intelligence” I argue for common limits based on perplexity and other measurements, with progress beyond them gated on our ability to understand AI behavior.
https://t.co/XX35m5vq89
$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.
Further reading: Here is Jakub Pachocki's full post on OpenAI's site, including the internal perplexity chart and his own list of four reasons the metric is still hard to use across labs.
https://t.co/PS5qcZs6XV
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?
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?
After 15 months of solo work, the author of the classic Mastering Emacs guide @mickeynp released a new AI coding workspace that is neither an Emacs plugin nor open source.
Even more paradoxical, the app itself is closed source, yet its core AI system instructions sit completely unlocked inside the settings menu for anyone to read and edit.
Most AI coding tools just tape a chat box beside a text editor, which falls apart when you try managing 15 to 50 AI agents simultaneously.
DreamCoder turns an ordinary web browser tab into a keyboard driven control room built specifically for supervising async machines.
> Mouse free navigation: Pressing Alt+W stamps instant letter labels onto every button, link, and data table on screen so you can jump anywhere or extract raw tables with 1 keystroke.
> Live subagent windows: When a main AI dispatches 2 background reviewers to inspect code, live status cards appear that let you pop open separate windows to watch each helper's reasoning in real time.
> Visual causal maps: Every prompt, tool action, and result links together in a zoomable flowchart alongside nested checklists that track clear completion goals.
> Zero config files: There are no sprawling settings files to maintain, and noisy terminal logs get trimmed by your own simple shell scripts before they can clog the AI's memory.
Even readers who normally hate web apps mimicking desktop window managers admitted this visual layout works surprisingly well, and the creator now uses DreamCoder exclusively to build DreamCoder itself.
Yet this old school craftsmanship sits beside a strict commercial license check that requires an online account even if you run local AI models offline on your own graphics card.
Do you actually want a visual control room to juggle 20 AI workers at once, or does reviewing that much machine output just turn coding into a full time auditing job?
@vitverb@mickeynp Purely for license enforcement. Because the free tier has zero feature limits up to 5 users, the app phones home to verify your license and disables even local models if that check fails.
If you want the full breakdown behind these fundamentals, including the gigawatt revenue model through 2031 and every moat and risk in detail, check out Drew Cohen's original piece here:
https://t.co/OFBOIF2wqz