Your ISP gave you a router.

It is not your router.

It can log every DNS query. Every device that connects. What time you go online. What time you stop.

On April 3, 2017, Trump signed S.J.Res 34 into law. In one signature, ISPs became legally allowed to sell your browsing history without your consent.

In the first half of 2024 alone, AT&T received 152,561 law enforcement demands for customer data. Verizon received 145,195. Straight from each company's own transparency report.

Then Comcast rents you the box. Xfinity xFi Gateway: $15 a month. $180 a year, for hardware you never own.

Then Netgear locks basic security behind Armor: $99.99 a year.

Then Amazon locks parental controls behind eero Plus: $9.99 a month.

Then most consumer router vendors drop firmware support within 3 to 5 years.

Then in December 2024, the U.S. Commerce Department opened an investigation into TP-Link, the most-shipped router brand in America, for potential national security ties to China.

You are renting access to your own internet in 2026.

Now meet OpenWrt.

A free and open source Linux operating system that flashes onto over 2,500 router models and gives you back full control of your own network.

Born in January 2004 after Linksys shipped the WRT54G running Linux firmware. Under the GPL, Linksys was required to publish the source. A small community took that code and rebuilt what a home router could be.

Twenty-two years later, OpenWrt runs on TP-Link. Netgear. Linksys. D-Link. Asus. GL.iNet. Ubiquiti. Xiaomi. Old routers whose vendors stopped patching still get fresh builds today.

27,681 stars. 12,693 forks. GPL-2.0.

- Full Linux root with 8,000+ packages
- WireGuard and OpenVPN server built in
- Ad blocking and DNS filtering network-wide
- Guest networks, VLANs, per-device firewall rules
- SQM traffic shaping for video calls
- Home Assistant, Prometheus, Grafana exporters
- LuCI web interface, no command line

Here is what OpenWrt costs:

Zero. Forever. On every device you flash.

Comcast charges $15 a month to rent a router. OpenWrt doesn't.
Netgear charges $99.99 a year for security. OpenWrt doesn't.
eero locks parental controls behind $9.99 a month. OpenWrt doesn't.

For one Xfinity subscriber, you save $180 a year.
For one Netgear Orbi owner, you save $99.99 a year.
For one eero household, you save $120 a year.

Your router. Your firmware. Your network.

100% Open Source. (Link in the comments)

Extracts license plate text and region from images using fast, lightweight OCR models supporting Keras 3 and ONNX.

https://github.com/ankandrew/fast-plate-ocr https://x.com/tom_doerr/status/2081952625684054285

You spent $2,000 on a MacBook Pro for a new hire.

Then $180 a year on the Microsoft 365 seat, $60 on Adobe, $144 on Slack, $240 on Zoom. All assigned to one laptop, one employee, one desk.

Then your CFO asked how many of the 400 laptops in this company are actually being used, and by whom.

You opened the Google Sheet. Half the rows were blank. Employees who left six months ago still had assets assigned.

So IT called ServiceNow. ITAM Standard is $100 per user per month. Professional $150. Enterprise $200+. The median ServiceNow contract on Vendr is $124,364 a year. Implementation runs three to five times the license fee.

Then IT called Lansweeper. $239 a month for 2,000 assets.

Then IT called Freshservice. $19 per agent, then metered on every asset over 500.

You are paying six figures a year to answer "who has laptop DELL-4471" in 2026.

Now meet Snipe-IT.

A free and open source IT asset management platform that runs on your own server and tracks every laptop, phone, license, and cable your company owns. No cloud. No agents. No per-seat fee.

Built in 2013 by a developer named Alison Gianotto who was the CTO of a New York City ad agency when the office had to move. Her team was tracking hundreds of laptops in a Google Doc. She tried every commercial and open source tool. None of them worked. So she wrote her own.

When she finished the first full inventory, she discovered the agency was missing $25,000 in assets.

Her husband nagged her for months to add a PayPal button. She finally did it. Two days later she had a paying customer. Snipe-IT itself stayed free.

14,113 stars. 3,886 forks. AGPL-3.0. Pushed to the repo today.

Here is what Snipe-IT gives you:

- Track every laptop, phone, monitor, cable, and accessory in one place
- Assign assets to users, departments, or locations with full history
- Software license management with seat tracking and expiration alerts
- Barcode and QR code generation for physical labels
- Custom fields, statuses, and categories
- LDAP, SAML, and Google Workspace SSO
- REST API for every endpoint
- Runs on any Linux box, Docker, or a $5 VPS

Here is what Snipe-IT costs:

Zero. Forever. No per-user fee. No asset cap.

ServiceNow charges $100 per user per month. Snipe-IT doesn't.
Lansweeper charges $239 a month for 2,000 assets. Snipe-IT doesn't.
Freshservice meters every asset over 500. Snipe-IT doesn't.
Ivanti and Flexera hide their pricing behind a sales call. Snipe-IT doesn't.

Here's the wildest part:

Snipe-IT is used by Sony, Yale University, Fujitsu, and thousands of school districts, hospitals, and government agencies that couldn't stomach a $124,000 ServiceNow bill. Alison's official title at Grokability is Chief Mohawk Officer.

For a 50-person IT team, you save $60,000 a year.
For a 500-person company, you save $600,000 a year.

Your assets. Your database. Your server.

100% Open Source. (Link in the comments)

A tricky LLM interview question:

You're serving a reasoning model on vLLM, and it keeps running out of GPU memory on long traces.

So you add KV cache compression and evict 90% of the cached tokens.

VRAM usage stays as is and GPU still runs out of memory.

Why?

(answer below)

Evicting 90% of the KV cache can free almost none of the memory it was using.

This sounds counterintuitive, but it follows directly from how production servers store the cache today.

The KV cache grows with every token a model generates. Each token appends its key and value vectors across every layer, and nothing is freed while generation continues.

This is the dominant memory cost for reasoning models.

If a 32K-token CoT caches ~32K tokens of KV vectors, a Qwen3-32B with 4-bit weights will run out-of-memory around 24K tokens on a 24GB GPU.

One obvious solution is to keep the important tokens and drop the rest, since attention is sparse enough to allow it.

But this does not solve the memory problem yet.

The reason is paged attention, which is the memory manager behind vLLM and most production servers.

Under the hood, it splits GPU memory into fixed physical blocks, each one holds the KV for about 16 tokens.

This block returns to the allocator only when every slot inside it is empty.

Since the eviction logic selects tokens by importance, and such tokens are scattered across blocks...

...so despite eviction, almost every block is left with at least some survivor tokens.

For instance, if the logic evicts 14k of 16k tokens across 1,000 blocks, most likely every block will still have a token.

This means the allocator frees almost nothing.

Placing the new tokens into those freed slots is not ideal because it breaks the cache's layout.

Say token 16,001 arrives, and it's placed in the slot the 40th token used to hold. The cache now reads position 38, then 16,001, then 41, so the cache is no longer in token order.

Attention can still compute the right answer from that, but only if every slot now carries a separate note recording which position it actually holds.

This introduces another bookkeeping cost that an in-order layout inherently avoids.

So the cache is logically 90% smaller and still physically the same size. Many compression results miss this because they measure on pre-allocated contiguous tensors rather than a paged server.

There's another problem.

Eviction methods pick which tokens to keep by looking at the attention scores themselves (as expected).

But fast attention kernels used in production, like FlashAttention, never save those scores.

They compute attention in small pieces and throw the full score grid away as they go, which is also why they're fast.

So the exact signal eviction methods need isn't available in memory. The workaround is to fall back to eager attention and build the full matrix, which gives up the speed FlashAttention was there to provide.

NVIDIA published a method called TriAttention to solve both these problems.

It never needs attention scores. Instead, it scores tokens from the geometry of the model's key and query vectors before RoPE is applied, where those vectors sit in stable clusters.

For the memory problem, it runs a compaction pass every 128 decoded tokens.

The surviving tokens slide forward to close the holes eviction creates, so whole blocks empty out and return to the allocator while the cache stays in token order.

On long reasoning traces, the approach matches full-attention accuracy while decoding 2.5x faster and using 10.7x less KV memory.

KV cache compression is a big infrastructure problem. The number that decides whether it works is the count of freed blocks, not the count of evicted tokens.

You can find the NVIDIA write-up here: https://research.nvidia.com/labs/eai/blogs/kv-cache-compression-and-its-infra-problems/

I wrote a first-principles breakdown of how the KV cache works. It walks through why the model stores keys and values at all, why the cache grows with every token, and a comparison of LLM generation speed with and without KV caching.

Read it below.

Photos & video were how the last generation expressed themselves.

Today’s form of expression is software.

Who will create the next viral game or app that plays right inside of the X Timeline?

Try the new Grok app builder. https://x.com/grok/status/2082134072793637196