local sampling + verification loops with zero token anxiety is the actual frontier.
just absolute unhinged autoresearch iteration without an API meter running in the back of your head.
we’re about to see a massive wave of optimizations that simply weren't legible before. the long wait is over.
spacy retweeted
There will be zero ability for humans to review any agent swarm work..it's just humanly not possible. We'll see a class divide - companies that trust AI and the ones that don't.
I've built harnesses that scale over massive data and semi broken human centric workflows. The only bottleneck left in enterprise now is human coordination. With agent swarms human managers will be left behind in the dust.
Here’s why agent swarms are going to be a very big deal. We can finally make compute a critical variable in a large number of important tasks in the economy which wasn’t possible before.
There’s a large amount of work in the world that is essentially only constrained by the amount of time that we put into research, data processing, discovery work, running tests, and more. All of this work is bottlenecked by our ability to plan, execute, review, and coordinate everything happening in these workflows. Which means we basically do very little of them today relative to their economic value.
Agent swarms are the first real ability to parallelize this work in a way where - in theory - the only parameter that matters is how much compute you throw at the problem. Cyber security, life sciences or material science research, trading, QA testing, data processing, and many other fields are essentially compute constrained problems in the limit once models are good enough at these tasks.
The breakthrough is that in a world where agents can use tools, write code, process data, have strong domain depth, and can coordinate with each other, you can begin to solve arbitrarily complex problems by setting the right goals and applying enough tokens to the problem. This changes a lot of fields as well as even the competitive dynamics in these fields.
the forward deployment engineer shortage is an illusion born from trying to make llms write raw sql or brittle tooling. graphjin.com shifts the paradigm back to the cpu where it belongs.
the real barrier to enterprise agi isn’t model intelligence, it’s the data integration tax. forcing llms to output sql or building brittle custom layers is completely unsustainable.
graphjin is a deterministic compiler harness that completely solves the enterprise data mismatch on the cpu
Notion AI lead @sarahmsachs argues that a lot of what runs on GPUs today should actually run on CPUs:
"We have a strong perspective at Notion that a lot of things should be served on CPU that are served on GPU. There’s a lot of benefit and determinism in executed code."
"Agents don’t need LLMs to reinvent that process every time, and that’s super token exhaustive, it’s super slow, and there’s no observability or determinism in it."
"Giving the agent a local file system to store local thoughts, to write short Python scripts, that’s really important when there’s so much information. That has been probably one of the most important tools outside of edit and search that our internal agents rely on."
@NotionHQ
love the vibe i need a desk with graffiti
Muse Spark Open -- weights releasing today -- runs great on new Mac Studios thanks to work with @johnternus and Apple.
nobody did it because they couldn’t. the legacy OEMs can barely make a touch screen entertainment system that doesn't lag 3 seconds, let alone solve vertical liftoff dynamics.
I'm not an Elon hater.
I'm just telling you the Honest Truth:
This is a marketing stunt.
The technology is quite old.
Just nobody ever thought of bolting it on a car.
Because it's pointless.
Porsche or Ferrari could also build something like this if they wanted, but they don't, because it's not worth it for them.
It's worth it for Elon though, because his entire brand is all about having the halo of futurism.
His marketing scales especially well with tech gimmicks like this, which is why he is financially incentivised to do them, while others aren't.
Again, I'm not hating, but actually educating yall.
This is not engineering genius.
It's marketing genius.
it is a literal crime that i haven’t posted about this yet.
the industry is wasting billions on model-side reasoning when the real bottleneck is just the deterministic boundary. if you are writing custom mcp tools or hand-crafting brittle text-to-sql prompts to connect an agent to a database, you are ngmi.
graphjin.com is the correct architecture here. it treats data, apis, and files as a single governed graph. it’s not an agent gateway; it’s a compiler harness. the agent reasons by emitting dynamic graphql, which the engine immediately compiles into a single, highly optimized query. zero n+1. zero resolver slop.
because the model interacts through the compiler harness rather than raw sql, you get total safety guaranteed at the engine level. 3,351 attempts on deeporg across 10 different models—including gemini-4-argon—and literally 0 unsafe effects.
stop letting llms do all the heavy lifting you don’t need a smarter prompt. you need a compiler.
spacy retweeted
Inspiration comes from the most unlikely places. Sometimes from movies that can reroute the future.
Illia Polosukhin was at Google in 2016, stuck on how machines should process language. He went to see Arrival. The squid like aliens heptapods don’t speak in sequence. They write in circular inkblots, each one a complete thought painted all at once. No beginning, no end. The entire meaning exists simultaneously.
That rewired something. What if instead of forcing a computer to read the way we do, one word after another, you let it look at the entire sentence, the entire paragraph, at the same time, and just decide which words were most relevant to each other? Even a word that appeared many paragraphs earlier might provide a contextual clue to what the next word meant.
That was the shift. Not just the sentence in front of you, but the whole text, held at once.
That idea became self-attention. Self-attention became the Transformer. And the Transformer is the architecture underneath every large language model that exists right now.
The thing that lets these models hold an entire context in mind, weigh relationships across thousands of words, and generate what comes next.
You all know that right?
movies are sorta dead everyone is watching youtube and streamers shot on iphones
Saying software is dead because everyone can create it with agents is like saying movies are dead because billions of ppl have a high res video camera in their pockets
Movies doing better than ever+more ppl make a living from videos, movies, TV than ever
(Analogy from @simonw)
those should be even easier as we can build deterministic validators for them while photoshop quality is subjective
Honestly: Photoshop and other creative apps are childs play.
The fun starts when the big guns get into focus: Highly technical CAD applications with license costs in the 5, 6 or 7 digits per seat per year.
Cadence Xcelium, Innovus and Virtuoso. Synopsys DesignCompiler, PrimeTime, Synplify. Siemens Questa. Altium. AutoCAD.
I recently built an open source Schematic Tracer that is objectively better than anything that Cadence or Synopsys have ever done. In 3 days.
They're all cooked.
people mourning the death of handwritten software is the ultimate status signaling trap.
have you actually interacted with human-written codebases out in the wild? 99% of legacy software is an unreadable tower of bugs, visual slop, and terrible UX held together by spit and duct tape. it’s pure junk optimized for corporate bloat cycles.
the romanticization of the manual programmer typing out every single character line-by-line is bizarre. we are replacing a fundamentally flawed human bottleneck with flawless, deterministically optimized intelligence loops. let’s go!
megacorps spent decades optimizing logistics loops to flood the world with low-margin mid slop.
generalist spatial intelligence completely flips the economic primitive. physical manufacturing bottlenecks vanish when a single local artisan can deploy zero-shot industrial automation from a garage.
we are completely unprepared for the sheer volume of high-quality indie brands about to delete big consumer legacy empires. the artisanal renaissance is a software update away.
the matter compiler is the ultimate primitives unlocked moment. if you think this is wild, you aren't ready for what happens when we chain-link multi-modal intelligence into dedicated hardware loops for 12 months straight.
we are actively accelerating toward the physical singularity. code in, anything out.
i fuckin love food trucks all restaurants should turn into food trucks
i love the food trucks it’s like food comes to you good food comes to you and what a great idea for breweries to collaborate with food trucks. it’s a win win literally this unknown brewery in Vancouver is so packed the beer is yum but the food is also a big draw.
also food trucks are forced to specialize they do few things but they do it good unlike a restaurant which has everything under the sun on their menu
tbh a few more hours from the model backed by gcp or aws it’ll be pretty solid too only issue i see is mail delivery maintaining clean ip’s etc
Yes, you can now build an inbox in an afternoon.
Congratulations, you are now responsible for:
- Deliverability.
- Bounce management.
- SPF, DKIM, DMARC.
- A domain that slowly drifts into spam.
- Finding out it why the hell it's drifting into spam?!
- Parsing emails from a mail server last updated in 2007
- HTML, plain text, and emails somehow containing neither.
- Outlook’s interpretation of HTML.
- Outlook’s other interpretation of HTML.
- Threading that breaks when someone changes the subject.
- Someone starting a new conversation by replying to an email from 2022.
- Forwards inside forwards inside forwards.
- CC, BCC, reply-all, and someone who really shouldn’t have clicked reply-all.
- Attachments.
- Attachments with malware..
- Inline images that become attachments.
- Attachments that were supposed to be inline images.
- A company logo attached to every reply, forever.
- Quoted replies.
- Distinguishing a quoted reply from the actual reply.
- Someone answering your five questions inside the quoted reply.
- Signatures longer than the conversation.
- The out-of-office that answers your out-of-office.
- Two autoresponders having a productive weekend together.
- Notifications that fire twice.
- Notifications that never fire.
- A reply sent from the notification itself.
- A reply to a notification about an internal note.
- Making absolutely sure that notes stays internal.
- Permissions.
- Two teammates writing the same reply.
- Drafts that sync across devices.
- Drafts that overwrite other drafts across devices.
- A confidently displayed “Saved” that was optimistic.
- Snoozed conversations.
- Time zones (!)
- Daylight saving time (!!)
- A conversation snoozed until a time that technically doesn’t exist.
- One customer showing up as three people across email, chat, and Slack.
- Three different people sharing one email address.
- Merging them.
- Undoing the merge.
- Search.
- Search that finds the thing you can literally see on screen.
- Webhooks arriving twice.
- Webhooks arriving out of order.
- Webhooks not arriving.
- Retrying a failed send without sending the customer six apologies.
- Rate limits you discover during your first traffic spike.
- Training your AI on your product.
- Scraping your website.
- Scraping the pages that need JavaScript.
- Not scraping your entire website footer into every answer.
- PDFs.
- PDFs that are actually pictures of PDFs.
- Keeping the content fresh.
- Knowing when content changed.
- Knowing when content disappeared.
- Making the AI forget the pricing page from last year.
- Conflicting answers across your docs, website, and past replies.
- Deciding which one wins.
- Chunking.
- Embeddings.
- Reindexing everything when you change embedding models.
- Retrieval that finds the answer, not just something with similar words.
- Web search tooling.
- Knowing when to search.
- Prompt injection hiding in a page you just scraped.
- An agent confidently promising a feature you don’t have.
- An agent confidently issuing a refund it shouldn’t.
- Tool permissions.
- Tool timeouts.
- Tool retries that issue the refund twice.
- Knowing when to hand off to a human.
- Actually handing off to a human.
- Not continuing to argue after handing off to a human.
- Evaluating whether any of this works.
- Evaluating it again after a model update.
- Token budgets.
- Context windows.
- A customer pasting an entire novel into the chat.
- WebSockets.
- WebSocket authentication.
- Reconnecting WebSockets.
- Reconnecting 10,000 WebSockets at once after a restart.
- A connection that looks alive but isn’t.
- A phone switching from Wi-Fi to mobile data mid-message.
- Messages that arrive once.
- Messages that arrive in order.
- Messages that arrive.
- Syncing read states.
- Unread badges.
- Typing indicators that stop typing eventually.
- A customer opening your widget in six tabs.
- Anonymous users becoming logged-in users.
- Logged-in users logging out on a shared computer.
- The iPhone keyboard covering the message composer.
- Safari.
- Keyboard navigation.
- A widget that must not slow down the actual product.
- Explaining that “sent” and “delivered” are different things.
- A help center.
- An editor.
- An editor that survives pasting from Google Docs.
- Images.
- Image resizing.
- Broken images.
- Article drafts.
- Autosave.
- Version history.
- Restoring the version before someone deleted half the article.
- Reordering nested categories without losing an afternoon.
- Slugs.
- 301 Redirects when slugs change.
- Internal links that now point nowhere.
- Working out which translations need updating after an edit.
- Search that understands what customers mean.
- Custom domains.
- DNS verification.
- SSL certificates.
- Renewing SSL certificates.
- Someone putting Cloudflare in front of your Cloudflare.
- Mobile layouts.
- Dark mode.
- Code blocks.
- Tables on small screens.
- Sitemaps.
- Canonical URLs.
- Accidentally indexing draft content.
- Accidentally deindexing published content.
- Cache invalidation.
- “I published it. Why does it still show the old version?”
- Making sure your AI sees the new version too.
- Spam.
- Phishing.
- Someone uploading malware as “a screenshot of the issue.”
- Someone pasting their CC details into chat.
- Someone asking you to delete every trace of that CC.
- Logs that also contain the CC.
- Backups that also contain the CC.
- Backups.
- Restoring backups.
- Discovering those are two separate features.
- An expired token.
- A background job that quietly stopped on Thursday.
- “Disk is full.” Sunday. 3 a.m.
- Monitoring all of this.
- Monitoring the monitoring.
Now you have two jobs 🤷♂️
