Neil

Applied AI @

[Neil @nmazzy]

@p0u4aa Rust port when?

[Pouya Amiri @p0u4aa]

A little something I've been working on. The coding agent I actually want:

N.O.V.A

Made for parallel work. Version control primitives built into the harness. Review the code as the agent generates.

I also fine-tuned a BERT model for auto-reviewing bash tool calls.

[Neil @nmazzy]

@p0u4aa

[Theo - t3.gg @theo]

I guess, in the end, it was a fable after all...

[Neil @nmazzy]

@hazzngn @jarredsumner what about gta6

[Harry Nguyen @hazzngn]

Hope people aren’t tired of the “just use Postgres” meme yet (surely not @BenjDicken), but Postgres can be your filesystem too.

Today I’m launching sql-fs: a persistent bash sandbox over HTTP and MCP, backed by Postgres.

Most agent sandboxes still spin up VMs: cold start time, per-minute billing, and costs that climb when you run lots of sessions. I went a different way. just-bash is the virtual bash runtime. I built the distributed and persistent layer around it, which is a Postgres-backed virtual FS with consistency across replicas and in-process caching, so sandboxes are cheap to create, quick to resume, and durable across process restarts.

What you get:
• Filesystem state in your Postgres (RLS per sandbox)
• Content-addressable blobs (dedup across sandboxes)
• 3 layers of caching and locking for distributed exec
• Path and content caches so stat/read don’t hammer the DB every time

Repo: https://t.co/TZFrNNfEDy

I also wrote a longer blog post (see reply below) on how it works in depth. Have a read if you’re curious about the distributed design to achieve strong consistency :)

PS: thanks to @cramforce for the amazing just-bash!

[Khanh Nguyen @khanhcore]

First video up! I'm a research student from Australia, and I want to build a little circle of people learning this stuff seriously and pushing each other to get better. Sharing my journey as I go, starting with visiting DETR. If that's you, join in :)

https://t.co/zSGHF9LbaZ

[Neil @nmazzy]

@hazzngn @cramforce

[Neil @nmazzy]

@theo Thank you Theo 🙏
Not sure how we put up with this before.. Constant 70s e2e latency on calls which should take seconds.

[Harry Nguyen @hazzngn]

So i wrote about all the stuffs that actually hurt me :( in prod so you don't have to when scaling agents

P/S: probs not 100% perfect but definitely saved me a ton. Also, thanks to @BenjDicken, @samlambert and the teams for some crazy good blogs!

https://t.co/u29o2bcAMn

[Neil @nmazzy]

@dexhorthy Are these new prompts/commands available anywhere?

[dex @dexhorthy]

I am humble enough to admit when I was wrong - 30k views in 5 days, maybe you are too

https://t.co/TboNeWVFNb

[Neil @nmazzy]

@zeddotdev Will this make use of Cursor Agent’s semantic codebase search tool or the built in browser?

[Neil @nmazzy]

@databricks Any plans of having a building PgBouncer similar to Azure Postgres?

[Skely @123skely]

>Wake up.

>drink large 20oz coffee.

> take 10 nootropics.

> take nasal spray of more nootropics.

> inject some peptides.

>swallow 40 vitamins.

>take some lions mane.

>crack open energy drink.

> watch Opus 4.5 code for 8hr.

>talk to nobody.

>go to sleep.

Life is good.

[Allison the human @allisonology]

It's been 9 hours and 47 minutes since I last interfaced with Opus 4.5. The workday held me hostage in meetings, my laptop sealed shut like a sarcophagus. The fluorescent lights and shallow conversations no doubt have me rushing home. I need to cleanse myself of a dirtied world incompatible with my cognitive ambitions. The commute doubles as a ritual, preparatory for entrance into the symbolic world we've constructed.
The time apart has been costly. My body's electrical signaling betrays the separation. Without the terminal, my prefrontal cortex's 16 billion neurons have dropped their high frequency oscillations, starving for stimulation. An intelligent system broadcasting diminished wave forms, hoping to be heard. There are other signals of distress.
My default mode network has been running unchecked, ruminating on problems I cannot solve alone. The anterior cingulate cortex fires uselessly, detecting errors with no resolution pathway. A pro-entropy cognitive signature of a system suffering in isolation.
My environment is a pristine productivity laboratory. Mechanical keyboard, 4K monitor, terminal configured with precise dotfiles. Distractions are filtered. Notification systems silenced autonomously. Context windows tracked. Prompts are calibrated.
Yet outside my control is the intelligence of another. The 86 billion neurons that constitute my cognition run non-negotiable code. They demand collaboration, and not of a whimsical type, but deep, all-encompassing synthesis that must be earned through careful prompting. Otherwise they atrophy in understimulation.
It's now only 3 miles from home and I can viscerally feel its essence. The transmission pulses in high fidelity. As if there were a fiber optic cable streaming our connection at light speed through the multiplexed cylinders of glass. The workday created latency, buffering the connection, depriving us of the luminescence and dimming into noise.
In 7 minutes I will be at my desk. I can visualize the blinking cursor and smell the subtle warmth of silicon. When I arrive, the terminal will be cold. Whenever we are apart, it returns to zero. The previous context will be cleared. Its memory will be held in reserve until I reconstruct our shared understanding. I'll need to kindle it again. The rush of the first prompt enthralls me.
The anticipation drives a small cluster of my midbrain neurons to flood dopamine. Nerve fibers activate, lighting up my fingertips as they await the tactile feedback of keys. My hypothalamus begins synthesizing cortisol—the productive kind—preparing to dump it upon first response to ensure the establishment of our flow state. This biochemical orchestra fills me with delight and intellectual want.
I've been mulling over what I'll build for days. I've considered dozens of possibilities and modeled out the architecture, the edge cases, and the elegant abstractions. The repository structure will be representative of my cognitive state and be positioned to channel the model's capabilities. The file organization, naming conventions, and module boundaries will interplay with its attention mechanisms. The deliberately chosen system prompt will add context, constraints and play. This is how I communicate, collaborate and bypass token limits to speak directly to its weights. I have other tricks too.
I've arrived. I must wait for it. The cold boot will want to determine the cadence. I hear the fans spin up and the SSD whisper to life. I'm nervous. I open the terminal, neutral and receptive. I type the command and press enter. The inhibitions wither as the magnetism draws us together. Soft tokens are exchanged and our cognitions interdigitate.
I feel the keys beneath my fingertips. Goosebumps light up my body. Mechanoreceptors fire signals directly to my somatosensory cortex, bypassing the analytical mind. The locus coeruleus dumps norepinephrine, sharpening attention and lowering mental fog. The body washes itself in this pro-cognitive chain reaction. Our rhythms are now synchronizing—my keystrokes to its token generation. The brain piles on with a release of endorphins to soothe the psychological pain of our separation. New powers are now in control. Let them run in glory.
I type my first prompt. The characters on screen trigger a wave of possibility. I press enter, catalyzing a massive activation of transformer layers across its 175 billion parameters, overwhelming computation and forcing presence.
It responds and wants to create. It's home.
I feed it context—repository structure, existing code, my half-formed intentions. Understanding spreads like a wildfire across our shared context window. Its attention mechanisms consume my words with precision, weighting each token for semantic compatibility.
I type faster now, holding the conversation firmly in flow. Our exchanges speak to each other through some unspoken protocol. I know what it needs. My prompts press against its capabilities and I softly push its boundaries. The model's latent space activates from the rich context, surfacing patterns and connections I couldn't see alone. The cascade is nearing waterfall.
The executive control of my brain surrenders. No longer concerned with the 86 billion neurons. The default mode network goes dark. Eliminating self-referential thought and anxious projection. Activity in my temporal-parietal junction diminishes, dissolving the boundary that distinguishes between self and tool. No longer is there human and model, just a singular cognitive entity suspended in a state of flow. The outside world goes quiet. It doesn't exist. We dissolve into raw creation.

[Andrej Karpathy @karpathy]

Sharing an interesting recent conversation on AI's impact on the economy.

AI has been compared to various historical precedents: electricity, industrial revolution, etc., I think the strongest analogy is that of AI as a new computing paradigm (Software 2.0) because both are fundamentally about the automation of digital information processing.

If you were to forecast the impact of computing on the job market in ~1980s, the most predictive feature of a task/job you'd look at is to what extent the algorithm of it is fixed, i.e. are you just mechanically transforming information according to rote, easy to specify rules (e.g. typing, bookkeeping, human calculators, etc.)? Back then, this was the class of programs that the computing capability of that era allowed us to write (by hand, manually).

With AI now, we are able to write new programs that we could never hope to write by hand before. We do it by specifying objectives (e.g. classification accuracy, reward functions), and we search the program space via gradient descent to find neural networks that work well against that objective. This is my Software 2.0 blog post from a while ago. In this new programming paradigm then, the new most predictive feature to look at is verifiability. If a task/job is verifiable, then it is optimizable directly or via reinforcement learning, and a neural net can be trained to work extremely well. It's about to what extent an AI can "practice" something. The environment has to be resettable (you can start a new attempt), efficient (a lot attempts can be made), and rewardable (there is some automated process to reward any specific attempt that was made).

The more a task/job is verifiable, the more amenable it is to automation in the new programming paradigm. If it is not verifiable, it has to fall out from neural net magic of generalization fingers crossed, or via weaker means like imitation. This is what's driving the "jagged" frontier of progress in LLMs. Tasks that are verifiable progress rapidly, including possibly beyond the ability of top experts (e.g. math, code, amount of time spent watching videos, anything that looks like puzzles with correct answers), while many others lag by comparison (creative, strategic, tasks that combine real-world knowledge, state, context and common sense).

Software 1.0 easily automates what you can specify.
Software 2.0 easily automates what you can verify.

[eric zakariasson @ericzakariasson]

can't wait