Trending

The daily dose, 11 Sep 2026

checked before you read it

Not everything that spreads on social media is true, so every claim below was checked against its primary source. The thread running through the day: Memory, proofs and job posts: the day the feed was mostly right, and wrong in the details.

12 items today, 29 claims checked against a primary source. 16 hold, 4 are partly right, 2 are wrong, 1 could not be verified and 6 are opinion.

one verdict per claim

Filters and days

Days

newest first

Kinds

Sep 11, 2026

Memory, proofs and job posts: the day the feed was mostly right, and wrong in the details.

12 items

29 claims · 16 hold · 4 partly · 2 wrong · 1 unverified · 6 opinion

Interview question

10 users can crash a 48 GB GPU running a 13B model. The maths says 60 GB.

An interview-style post: a 13-billion-parameter model is deployed on a 48 GB GPU, only 10 users are on it, and the server still dies with CUDA out of memory. The caption blames the KV cache, uncontrolled concurrency and memory fragmentation, and prescribes continuous batching (vLLM, TensorRT-LLM, TGI), capping tokens and context, keeping 10 to 15 percent of memory free, and queueing with backpressure.

26 GB13B weights, fp16
0.8 MBKV per token, 13B
60 GB10 users × 4k tokens

A 13B model needs about 26 GB in fp16, so 48 GB looks like plenty.13 billion parameters at 2 bytes each is 26 GB before anything else is loaded. The weights are the smallest part of the surprise.
The KV cache is what actually fills the card.For Llama-2-13B (40 layers, hidden size 5,120) every token keeps 2 × 40 × 5,120 × 2 bytes ≈ 0.8 MB of keys and values. A 4,096-token conversation is 3.4 GB. Ten of them are 34 GB. Add 26 GB of weights and you are at 60 GB on a 48 GB card.Llama 2: Open Foundation and Fine-Tuned Chat Models (Table 1)
KV growth is exponential.It is linear: tokens × concurrent sequences × a fixed per-token cost. Linear is enough to kill you, but calling it exponential leads people to the wrong fixes.Efficient Memory Management for Large Language Model Serving with PagedAttention
Fragmentation means a free 15 GB may have no 3 GB block to give.The vLLM paper measured 60 to 80 percent of KV memory wasted by fragmentation and over-reservation in earlier servers. PagedAttention fixes it by storing the cache in fixed-size blocks, the way an operating system pages memory.Efficient Memory Management for Large Language Model Serving with PagedAttention
Set gpu_memory_utilization=0.85 in vLLM.vLLM's default is 0.9; the post's 0.85 is a reasonable choice, not the setting that prevents the crash. What prevents it is that vLLM pre-allocates the KV budget inside that fraction and admits only as many sequences as fit.vLLM engine arguments

the line to remember

Serving memory = weights + (tokens × concurrent users × per-token KV cost). Size the card for the cache, not the model.

For your product

If a vendor quotes you a GPU from the model size alone, the quote is wrong. Ask for the per-token KV cost, the context cap and the concurrency cap; those three numbers decide whether the box is enough.

Go deeperKV cache Batching and PagedAttention Prefill and decode Quantisation Model gateway architecture

Sources: Llama 2: Open Foundation and Fine-Tuned Chat Models (Meta) · Efficient Memory Management for LLM Serving with PagedAttention (UC Berkeley (vLLM)) · vLLM engine arguments (vLLM)

Research

11 days, 13 million lines of Lean: Claude formalised Fermat's Last Theorem.

A post claims Claude formalised Fermat's Last Theorem in eleven days, largely autonomously, and that nobody will ever read the proof. Numbers on screen: 30,300, 29,500 and about 6 billion.

11 dayslargely autonomous
29,500theorems in the proof
~6Boutput tokens

the line to remember

Long-running agent swarms with a hard verifier (here, Lean's kernel) can finish multi-year work in days. The verifier is what makes autonomy safe.

For your product

Where your work has a checker (tests, a compiler, a ledger that must balance, a schema), agents can run for days unattended. Where it does not, they cannot. Build the checker first.

Go deeperThe agent loop Evals Tool calling Evaluation loop architecture

Sources: Formalizing Fermat's Last Theorem (Anthropic) · Anthropic AI 'formalizes' proof of Fermat's last theorem (Nature)

Interview question

All weights set to 0: every neuron gets the same gradient, so the network never learns.

A whiteboard post asks whether you can initialise all weights to zero. Answer: no, the gradient is identical for every neuron, so weights must be random to break symmetry. A commenter adds that with no bias term everything stays exactly zero; another says use Glorot for sigmoid or tanh and He for ReLU.

the line to remember

Initialisation is about breaking symmetry and keeping signal variance stable through the layers. Zero does neither.

For your product

Nothing to decide here unless you train models. If you do, this is the first question a reviewer will ask about any training bug that ends with a network that outputs the same answer for everything.

Go deeperNeural networks Backpropagation Optimisers: SGD, momentum and Adam

Sources: Understanding the difficulty of training deep feedforward neural networks (Glorot and Bengio, AISTATS) · Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification (He et al.)

Concept

3 agent patterns on one graphic: CodeAct, ReAct and agentic RAG. Two of the three descriptions oversell.

A graphic contrasts a single agent with a multi-agent system, then defines CodeAct (the agent acts by writing and running Python), ReAct (reasoning traces interleaved with tool actions, said to overcome hallucination and error propagation) and agentic RAG (agents orchestrating the retrieval pipeline).

the line to remember

ReAct is the loop, CodeAct is the action language, agentic RAG is the loop pointed at your documents. Pick the simplest one that fits the task.

For your product

Most business tasks do not need a multi-agent system. A single agent with a few well-described tools and a stop condition covers a lot; add agents only when the path cannot be known in advance.

Go deeperThe agent loop Tool calling Retrieval-augmented generation (RAG)Agent memory Single agent with tools Agentic graph RAG

Sources: ReAct: Synergizing Reasoning and Acting in Language Models (Yao et al., Princeton and Google) · Executable Code Actions Elicit Better LLM Agents (Wang et al., UIUC)

Research

91.2 percent: how often innocent tool calls could be chained into a harmful action in Amazon's STAC study.

Two papers led by Amazon interns were accepted: one on relational priors in LLM multi-agent systems (AACL) and STAC, on how benign tools can form dangerous chains for LLM agents (EMNLP REALM workshop).

91.2%mean final attack success (STAC)
483generated attack chains

the line to remember

Agreement between agents is not accuracy, and per-tool safety is not chain safety. Evaluate the sequence, not the pieces.

For your product

If your agent can read, write and send, the danger is the combination. Approval gates on the final effect (money moved, mail sent, record deleted) matter more than filters on each tool.

Go deeperPrompt injection Guardrails The agent loop Model Context Protocol (MCP)

Sources: Relational Priors as Convergence Pressure in LLM-Based Multi-Agent Systems (Shen et al., Amazon) · STAC: When Innocent Tools Form Dangerous Chains for LLM Agents (Li et al., Amazon)

Tools

8 steps and 12 products on one graphic. The model names check out; the integration counts do not.

A graphic lays out building an agent in eight steps (purpose and scope, system prompt, choose the LLM, tools and MCP servers, memory, orchestration, interface, testing and evals) and tabulates products: ChatGPT on GPT-5.6, Claude on Opus 5, Perplexity, Cursor, Windsurf on SWE-1.6, Claude Code on Claude 5, Lindy, Zapier Agents, n8n, LangGraph, CrewAI, LlamaIndex.

the line to remember

The stack in the table is real. The step that decides the outcome, evals, is listed last and should come first.

For your product

When a vendor shows you a build plan, look for where evaluation sits. If it is the final box, ask how they will know the agent works before that point.

Go deeperThe agent loop Agent memory Tool calling Model Context Protocol (MCP)Evals How we work: the six stages

Sources: GPT-5.6: Frontier intelligence that scales with your ambition (OpenAI) · Introducing SWE 1.6: Improving Model UX (Cognition)

Viral claim

The Navier–Stokes equations are not 'one of the hardest ever solved'. They are unsolved, and AI just found new singularities.

A maths page introduces the Navier–Stokes equations, credits 'C. Navier, G. Stokes, 1845', says they predict weather and how planes fly, and calls them still one of the hardest equations ever solved. A separate post shows a vortex with 'inward spiral' and 'axial stretching' labels.

the line to remember

Neural networks trained on equations rather than data can find solutions humans could not tune by hand. The proof still has to be done by mathematicians.

For your product

The same tool family (networks fitted to physics, not to examples) is what makes AI useful in simulation-heavy businesses: aerodynamics, weather, materials, chip cooling. If your work has equations, there is a model for it.

Go deeperNeural networks Loss and gradient descent Optimisers: SGD, momentum and Adam

Sources: Navier–Stokes Equation (Clay Mathematics Institute) · Discovery of Unstable Singularities (Wang et al., Google DeepMind and universities) · Discovering new solutions to century-old problems in fluid dynamics (Google DeepMind)

Concept

32 parallel paths beat one wider block: ResNeXt's 'cardinality', doing the rounds again in 2026.

A walkthrough of 'Aggregated Residual Transformations for Deep Neural Networks' (Xie, Girshick, Dollár, Tu, He), the paper that introduced cardinality: repeating a block that aggregates a set of transformations with the same topology.

the line to remember

Split, transform, merge. The same idea now runs in mixture-of-experts language models: many small parallel paths instead of one big one.

For your product

A 2016 vision paper is still on the feed because the idea generalised. Grouped, parallel computation is why today's largest models can be cheap per token.

Go deeperConvolution and vision Residual stream and normalisation Mixture of experts

Sources: Aggregated Residual Transformations for Deep Neural Networks (Xie et al., UC San Diego and Facebook AI Research, CVPR)

Hiring signal

Billions of images and videos a month: Adobe is hiring for the data pipeline behind Firefly, not the model.

A job post for staff or principal roles: the team scales the data pipeline to billions of images and videos per month and needs someone who can make VLM and LLM inference fly, with experience curating data for image and video diffusion models at scale.

the line to remember

Frontier labs spend as much engineering on inference for data curation as on training. The unglamorous pipeline is where the senior roles are.

For your product

The same skill, running vision-language models cheaply over a large archive, is what turns your document, image or video backlog into searchable, structured data. It is a solved engineering problem at Adobe's scale and a small project at yours.

Go deeperDiffusion models Prefill and decode Speculative decoding Batching and PagedAttention Visual document retrieval architecture

Sources: Seen on the feed; no factual claims to source

Hiring signal

Thousands of GPUs, petabytes per run, no ML experience required: Mistral's pretraining team wants systems engineers.

Mistral is hiring for its pretraining team: fault-tolerant training across thousands of GPUs, data-loading systems that process petabytes of multimodal data per run, and a training codebase used by hundreds of people. Prior ML experience is not required; distributed systems, performance and data pipelines are.

the line to remember

The model is a few thousand lines; the training system is the product. Reliability and throughput engineers are the scarce hires.

For your product

You will almost never pretrain. But the same reliability habits (checkpointing, retries, idempotent pipelines, throughput budgets) are what keep a fine-tuning or batch-inference job from burning a week of cloud spend when a node dies.

Go deeperPretraining and scaling laws Distillation Fine-tuning and LoRA

Sources: Seen on the feed; no factual claims to source

Hiring signal

One job post, five nouns: data pipelines, RAGs, MCPs, evals and human approval steps.

An AI engineer role: turn client workflows into working AI solutions from prototype to deployment; build assistants, RAG systems and automated workflows; design evals for answer quality, task completion and failure cases; implement access controls, human approval steps and error handling; monitor reliability, latency and cost; document and hand over. Asks for LangChain, LangGraph, Snowflake and an agent orchestration tool, and for the ability to explain systems to non-technical stakeholders. Apply with one or two projects, explaining the problem, what you built and how you evaluated it.

the line to remember

The market's definition of an AI engineer in 2026: someone who can build the workflow and prove it works, then hand it over.

For your product

Use this post as an interview checklist for any vendor: ask to see their eval set, their approval gates and their handover documentation before you ask to see a demo.

Go deeperRetrieval-augmented generation (RAG)Model Context Protocol (MCP)Evals Guardrails LLM observability How we work No lock-in, full documentation

Sources: Seen on the feed; no factual claims to source

Roadmap

40 system design problems in 8 groups, from 'design Twitter' to the boss fights. Now a track on our roadmap.

A shared list of the classic interview problems grouped into social feeds, messaging, e-commerce and payments, core infrastructure, events and queues, analytics and logging, IDs and coordination, and the boss fights (Uber, Ticketmaster, Google Search, a trading platform, a container orchestrator).

the line to remember

Every AI system we build is also one of these: a feed, a queue, a cache, a search index, an ID generator. The AI part is the smaller half.

For your product

When an AI feature is slow or expensive, the cause is usually on this list (caching, queues, fan-out, storage) rather than the model. Ask your vendor which of these your system contains.

Go deeperHNSW vector search BM25 keyword search Hybrid search and reranking KV cache System design track on the roadmap Data engineering and cloud

Sources: Seen on the feed; no factual claims to source

Want a second pair of eyes on what you are being sold?

Send us the deck, the demo or the post. We check the claims the same way and tell you what it means for your product.

Book a discovery call Send us a brief