Lgoswami

Fall seven times, stand up eight. True strength lies not in never falling, but in rising every time you do.

[ani @anirudhbv_ce]

I implemented @GoogleResearch's TurboQuant as a CUDA-native compression engine on Blackwell B200.

5x KV cache compression on Qwen 2.5-1.5B, near-loseless attention scores, generating live from compressed memory.

5 custom cuTile CUDA kernels ft:
- fused attention (with QJL corrections)
- online softmax
-on-chip cache decompression
- pipelined TMA loads

Try it out: https://t.co/m5vkJxWIY6

s/o @blelbach and the cuTile team at @nvidia for lending me Blackwell GPU access :)

cc @sundeep @GavinSherry

[Jacob Coxon @hilbertspaess]

I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives. More thoughts below.

[Sebastian Raschka @rasbt]

Reasoning from scratch round 2: In this video, I cover the text generation process in LLMs and KV caching (to prepare the base model before adding reasoning techniques in the upcoming ones).

00:00 Introduction and reasoning model demo
01:55 How to work through the book
05:00 Chapter 2 overview
08:25 Checking PyTorch and hardware support
10:26 Apple silicon and MPS caveats
15:00 Cloud GPU options
16:08 Tokens and tokenization
18:20 Qwen3 and the Reasoning From Scratch package
23:05 Encoding and decoding text
26:24 Downloading weights and selecting a device
31:01 Loading the pretrained Qwen3 model
34:32 How LLMs generate text
36:47 Input tensors and batch dimensions
41:48 Running the model in inference mode
44:11 Logits and next-token predictions
49:21 Greedy decoding with argmax
52:28 Building a streaming text generator
01:01:28 Generating text and handling end-of-sequence tokens
01:06:00 Benchmarking text generation
01:14:34 How KV caching works
01:17:22 Adding KV caching and measuring the speedup
01:24:31 Model compilation with torch.compile
01:30:33 Combining compilation with KV caching
01:32:53 Comparing CPU and GPU performance
01:35:32 Recap and next steps

[Christian Flanders @CFlanders7]

Same is true for trading. Amateurs are in for the excitement

[Joshua Saxe @joshua_saxe]

Finally listened to this Ajeya Cotra interview and it's very good. Security friends: misalignment risk is not a conspiracy or a marketing stunt. Folks who care about practical responsible ML issues: agent swarms are not a fairy tale. I've updated in the last month; I had seen loss of control, scheming, and reward hacking 2023-2025 as worthwhile academic research but impractical and had suspected these topics might turn out to be as marginal as the adversarial example lit was to security from the 2010s. My update is that these all these risks are now clearly extremely practical and I care way more about them now and I think security folks should too, because solving them will include security skillsets

[Sam Lambert @samlambert]

We've shipped a ridiculous amount of technical blog posts in the last month.

Read below for the full list.

[ClaudeDevs @ClaudeDevs]

Claude Fable 5 changed how we work on the Claude Code team day to day.

We used to verify that Claude did the work right. Now we verify that it's doing the right work.

Here’s the 3 biggest changes:

[Dan Robinson @danrobinson]

Natural language autoencoders are the discovery that most makes me question my research intuition

You jointly train two models: one to turn activations into English, and one to turn English into activations.

The objective for the second model is to invert the first one, and the objective for the first model is... to be invertible by the second one

I would not have expected this to work as well as it seems to!

[Anthropic @AnthropicAI]

New Anthropic research: Natural Language Autoencoders.

Models like Claude talk in words but think in numbers. The numbers—called activations—encode Claude’s thoughts, but not in a language we can read.

Here, we train Claude to translate its activations into human-readable text.

[rari @0xwhrrari]

Anthropic hired this engineer at $250K-$750K a year because he knows how to build harnesses for multi-agent systems

In this 15-minute workshop, he shows exactly how to build one from scratch

AI → Agents → Harness → Loops → Graphs

step 1 → start with the Claude Agent SDK - the harness handles loops, context, and sandboxing

step 2 → separate the brain from the hands - reasoning in one place, tools in a sandbox, 60% faster to first token

step 3 → run it server-side and log every step - close your laptop and it keeps running, crashes resume from the log

step 4 → make failure cheap - retry dead sandboxes and replay lost context instead of starting over

step 5 → turn yesterday's logs into new memory and skills - the harness wakes up smarter

Anthropic calls this "dreaming"

Most people spend weeks building this by hand

You don't have to

Bookmark and watch it

Then read the full harness engineering guide below ↓

[Dongxi 东锡 NLP @dongxi_nlp]

推荐阅读！

GPT-6 Astra + Recurrent depth /Looped transformer

[Ivan Velichko @iximiuz]

What Exactly Does "docker run -it" Do? 🧐

When a container needs to run an interactive app such as a shell, language REPL, or text editor, the "docker run" command requires two extra flags: -i and -t.

Explore how they work by solving this challenge: https://t.co/x7IPGkvPI1

[Ahmad @TheAhmadOsman]

Some ideas about how to get started learning about all of this

Inference Engines & Topics

- vLLM: PagedAttention, continuous batching, prefix caching, CUDA graphs

- SGLang: RadixAttention/prefix reuse, speculative decoding, MoE, structured/agent workloads

- TensorRT-LLM: NVIDIA peak stack, FP8/FP4, Wide-EP, disaggregated serving

- FlashInfer: reusable kernel/operator library for attention/GEMM/MoE/sampling

Kernels

- Triton tutorials → custom fused kernels

- CUTLASS/CuTe → Tensor Core GEMM and Blackwell/Hopper details

- FlashAttention papers → attention algorithm/kernel co-design

- PagedAttention paper → KV-cache memory management

- MoE docs → routing + grouped GEMM + all-to-all

- Nsight profiling → stop guessing

Do this mini-project sequence

1. Implement RMSNorm in Triton; compare to PyTorch

2. Implement fused SiLU × gate

3. Implement simple FP16 matmul; compare to cuBLAS/rocBLAS

4. Implement paged KV lookup for decode attention

5. Add FP8 KV cache with per-block scales

6. Implement toy top-k sampling on GPU

7. Implement tiny MoE dispatch + grouped GEMM

8. Integrate one custom op into vLLM or SGLang and profile end-to-end

[Josh Long @starbuxman]

Amazing

[Immanuel @immanuel_vibe]

eBPF is genuinely wild and I don't think enough people have clocked it yet

the trick is XDP, it runs down at the driver layer, so you drop junk packets before the kernel even builds an skb.
Cloudflare kills millions of malicious packets per second this way.
the network stack never spends a cycle on them

Netflix traces flow logs across their entire fleet with it.
No tcpdump melting the CPU
they also caught a noisy-neighbor disk latency bug that normal tools were blind to, because the latency was hiding between the syscall and the disk

Google's GKE dataplane v2 is Cilium/eBPF.
they've written about running it on 65,000-node clusters, such a stupid-big number

and you don't hand-write BPF bytecode like it's 2016, you just install bpftrace and run a one-liner
a live histogram of your prod read sizes - no recompile, no restart, no downtime

and it can't crash the box.
it's a tiny sandboxed VM inside the kernel, and the verifier rejects anything unsafe before it ever loads

observability without the observer effect, finally

also btw I'm building pktz because of all this. eBPF network monitor, per process, per connection, live.
https://t.co/SpTi3U8Ny2

what's the one bpftrace line you keep in your back pocket?

[Raul Junco @RaulJuncoV]

After years building event-driven systems.

Here are the top 4 mistakes I have seen:

1. Duplication

Events often get re-delivered due to retries or system failures. Without proper handling, duplicate events can:

• Charge a customer twice for the same transaction.
• Cause duplicate inventory updates, messing up stock levels.
• Create inconsistent or broken system states.

Solution:

• Assign unique IDs to every event so consumers can track and ignore duplicates.
• Design event processing to be idempotent, ensuring repeated actions don’t cause harm.

2. Not Guaranteeing Order

Events can arrive out of order when distributed across partitions or queues. This can lead to:

• Processing a refund before the payment.
• Breaking logic that relies on correct sequence.

Solution:

• Use brokers that support ordering guarantees (e.g., Kafka).
• Add sequence numbers or timestamps to events so consumers can detect and reorder them if needed.

3. The Dual Write Problem

When writing to a database and publishing an event, one might succeed while the other fails. This can:

• Lose events, leaving downstream systems uninformed.
• Cause mismatched states between the database and event consumers.

Solution:

• Use the Transactional Outbox Pattern: Store events in the database as part of the same transaction, then publish them separately.
• Adopt Change Data Capture (CDC) tools to track and publish database changes as events automatically.

4. Non-Backward-Compatible Changes

Changing event schemas without considering existing consumers can break systems. For example:

• Removing a field might cause missing data for consumers.
• Renaming or changing field types can trigger runtime errors.

Solution:

• Maintain versioned schemas to allow smooth migration for consumers.
• Use formats like Avro or Protobuf that support schema evolution.
• Add adapters to translate new schema versions into older ones for compatibility.

Every schema change is a test of your system’s resilience.

What other mistakes have you seen out there?

[Diyi Yang @Diyi_Yang]

This fall @michaelryan207 @jyangballin and I are teaching a new course CS329Z "Engineering AI Agents" on how to build AI Agents from scratch. Come join us and learn how to build them 🤖

[Hriday Kadam @hridaykadam]

people doing just leetcode - RIP

[Immanuel @immanuel_vibe]

you don't really understand AWS networking until none of these surprise you:

1. a subnet isn't "public" because of a checkbox.
it's public because its route table has a route to an internet gateway. that's the entire definition

2. security groups are stateful.
NACLs are stateless - forget to allow the ephemeral port range and your return traffic dies silently

3. security groups attach to ENIs, not instances.
and they can reference other security groups instead of CIDRs.
stop hardcoding IPs

4. NAT Gateways are per-AZ.
one NAT in one AZ = a cross-AZ data charge on every byte, plus a single point of failure you're paying a premium for

5. S3 and DynamoDB gateway endpoints are free.
if your NAT bill makes you flinch, that's usually the reason

6. AWS reserves 5 IPs in every subnet. A /28 gives you 11 usable hosts, not 16

7. VPC peering is not transitive, and it won't route through to the peer's internet gateway.
It's a cable between two VPCs, nothing more

8. cross-AZ traffic costs money in both directions - even between private IPs, even behind a load balancer

9. if you want to know more about AWS networking - https://t.co/V2Z6hHXZLh

which one bit you in prod?

[seeco @seeconvm]

Anthropic engineer at Stanford:

"80% of our engineers are using self-improving loops"

"Now everyone is building agentic graphs"

In a 1-hour lecture at Stanford, an Anthropic engineer reveals how Claude actually thinks

And why traditional single-turn prompting is officially dead

Linear agent workflows break, directed graphs are replacing them

Single-turn prompts fail on real-world projects, graphs let agents reason, test, fail, and self-correct until the job is done

This 1-hour watch will literally replace your $500 course on agentic engineering

Watch it today

Then read how to become a Graph Engineer in the article below