{"version":1,"site":{"name":"Tachyon","url":"https://tachyon.techntrek.in"},"feed":{"title":"Tachyon: the daily dose","url":"https://tachyon.techntrek.in/trending","self":"https://tachyon.techntrek.in/feed/trending","rss":"https://tachyon.techntrek.in/feed/rss"},"generatedAt":"2026-09-11T20:05:35.713Z","days":[{"date":"2026-09-11","url":"https://tachyon.techntrek.in/trending/2026-09-11","theme":"Memory, proofs and job posts: the day the feed was mostly right, and wrong in the details.","itemCount":12,"verdicts":{"holds":16,"partly":4,"wrong":2,"unverified":1,"opinion":6},"items":[{"slug":"cuda-oom-48gb-ten-users","url":"https://tachyon.techntrek.in/trending/2026-09-11#cuda-oom-48gb-ten-users","date":"2026-09-11","kind":"interview","title":"10 users can crash a 48 GB GPU running a 13B model. The maths says 60 GB.","said":"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.","takeaway":"Serving memory = weights + (tokens × concurrent users × per-token KV cost). Size the card for the cache, not the model.","forBusiness":"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.","numbers":[{"value":"26 GB","label":"13B weights, fp16"},{"value":"0.8 MB","label":"KV per token, 13B"},{"value":"60 GB","label":"10 users × 4k tokens"}],"verdicts":{"holds":3,"partly":1,"wrong":1,"unverified":0,"opinion":0},"checks":[{"claim":"A 13B model needs about 26 GB in fp16, so 48 GB looks like plenty.","verdict":"holds","note":"13 billion parameters at 2 bytes each is 26 GB before anything else is loaded. The weights are the smallest part of the surprise."},{"claim":"The KV cache is what actually fills the card.","verdict":"holds","note":"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.","source":{"title":"Llama 2: Open Foundation and Fine-Tuned Chat Models (Table 1)","url":"https://arxiv.org/abs/2307.09288"}},{"claim":"KV growth is exponential.","verdict":"wrong","note":"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.","source":{"title":"Efficient Memory Management for Large Language Model Serving with PagedAttention","url":"https://arxiv.org/abs/2309.06180"}},{"claim":"Fragmentation means a free 15 GB may have no 3 GB block to give.","verdict":"holds","note":"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.","source":{"title":"Efficient Memory Management for Large Language Model Serving with PagedAttention","url":"https://arxiv.org/abs/2309.06180"}},{"claim":"Set gpu_memory_utilization=0.85 in vLLM.","verdict":"partly","note":"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.","source":{"title":"vLLM engine arguments","url":"https://docs.vllm.ai/en/latest/configuration/engine_args/"}}],"learn":[{"slug":"kv-cache","title":"KV cache","url":"https://tachyon.techntrek.in/learn/kv-cache"},{"slug":"batching-and-paged-attention","title":"Batching and PagedAttention","url":"https://tachyon.techntrek.in/learn/batching-and-paged-attention"},{"slug":"prefill-and-decode","title":"Prefill and decode","url":"https://tachyon.techntrek.in/learn/prefill-and-decode"},{"slug":"quantization","title":"Quantisation","url":"https://tachyon.techntrek.in/learn/quantization"}],"read":[{"label":"Model gateway architecture","url":"https://tachyon.techntrek.in/architectures/model-gateway"}],"sources":[{"title":"Llama 2: Open Foundation and Fine-Tuned Chat Models","url":"https://arxiv.org/abs/2307.09288","org":"Meta","date":"2023"},{"title":"Efficient Memory Management for LLM Serving with PagedAttention","url":"https://arxiv.org/abs/2309.06180","org":"UC Berkeley (vLLM)","date":"2023"},{"title":"vLLM engine arguments","url":"https://docs.vllm.ai/en/latest/configuration/engine_args/","org":"vLLM","date":"2026"}],"tags":["inference","GPU","KV cache","interview"]},{"slug":"claude-formalised-fermat","url":"https://tachyon.techntrek.in/trending/2026-09-11#claude-formalised-fermat","date":"2026-09-11","kind":"research","title":"11 days, 13 million lines of Lean: Claude formalised Fermat's Last Theorem.","said":"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.","takeaway":"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.","forBusiness":"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.","numbers":[{"value":"11 days","label":"largely autonomous"},{"value":"29,500","label":"theorems in the proof"},{"value":"~6B","label":"output tokens"}],"verdicts":{"holds":2,"partly":0,"wrong":0,"unverified":0,"opinion":1},"checks":[{"claim":"Claude produced a computer-checked proof of Fermat's Last Theorem in 11 days.","verdict":"holds","note":"Anthropic's research post reports the first end-to-end formalisation of the theorem in Lean, produced in 11 days by dozens of Claude agents running a multi-agent workflow. Nature covered it as a milestone for mathematics.","source":{"title":"Formalizing Fermat's Last Theorem","url":"https://www.anthropic.com/research/formalizing-fermats-last-theorem"}},{"claim":"The numbers on screen: 30,300 theorems, 29,500 used, about 6 billion.","verdict":"holds","note":"The run proved 30,300 intermediate theorems, 29,500 of which sit in the final proof, wrote about 13 million lines of Lean and used roughly 6 billion output tokens.","source":{"title":"Formalizing Fermat's Last Theorem","url":"https://www.anthropic.com/research/formalizing-fermats-last-theorem"}},{"claim":"Nobody will ever read the proof.","verdict":"opinion","note":"That is the point of a proof assistant: Lean's small kernel checks every step, so humans read the statement and the kernel's verdict, not 13 million lines. Kevin Buzzard's human-led formalisation project, started in 2024, was expected to take years; the AI run compresses that, it does not replace Andrew Wiles's 1995 proof."}],"learn":[{"slug":"agent-loop","title":"The agent loop","url":"https://tachyon.techntrek.in/learn/agent-loop"},{"slug":"evals","title":"Evals","url":"https://tachyon.techntrek.in/learn/evals"},{"slug":"tool-calling","title":"Tool calling","url":"https://tachyon.techntrek.in/learn/tool-calling"}],"read":[{"label":"Evaluation loop architecture","url":"https://tachyon.techntrek.in/architectures/evaluation-loop"}],"sources":[{"title":"Formalizing Fermat's Last Theorem","url":"https://www.anthropic.com/research/formalizing-fermats-last-theorem","org":"Anthropic","date":"2026"},{"title":"Anthropic AI 'formalizes' proof of Fermat's last theorem","url":"https://www.nature.com/articles/d41586-026-02822-9","org":"Nature","date":"2026"}],"tags":["agents","maths","Anthropic","Lean"]},{"slug":"zero-init-weights-interview","url":"https://tachyon.techntrek.in/trending/2026-09-11#zero-init-weights-interview","date":"2026-09-11","kind":"interview","title":"All weights set to 0: every neuron gets the same gradient, so the network never learns.","said":"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.","takeaway":"Initialisation is about breaking symmetry and keeping signal variance stable through the layers. Zero does neither.","forBusiness":"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.","numbers":[],"verdicts":{"holds":2,"partly":1,"wrong":0,"unverified":0,"opinion":0},"checks":[{"claim":"With all weights equal, every neuron in a layer receives the same gradient.","verdict":"holds","note":"Identical weights give identical activations and identical error signals, so each update is the same for every neuron in the layer. They stay clones forever. Random initialisation breaks the tie.","source":{"title":"Understanding the difficulty of training deep feedforward neural networks","url":"https://proceedings.mlr.press/v9/glorot10a.html"}},{"claim":"With no bias, everything will be zero.","verdict":"partly","note":"True for ReLU and tanh: zero weights give zero hidden activations, so every weight gradient except the output bias is zero and nothing moves. With sigmoid the hidden activations are 0.5, so the output layer does learn, but the hidden neurons remain identical. The post and the comment describe different activation functions."},{"claim":"Use Glorot for sigmoid or tanh, He for ReLU.","verdict":"holds","note":"Glorot and Bengio (2010) scale the initial variance by fan-in and fan-out for symmetric activations; He et al. (2015) double it for ReLU, which zeroes half its inputs.","source":{"title":"Delving Deep into Rectifiers","url":"https://arxiv.org/abs/1502.01852"}}],"learn":[{"slug":"neural-networks","title":"Neural networks","url":"https://tachyon.techntrek.in/learn/neural-networks"},{"slug":"backpropagation","title":"Backpropagation","url":"https://tachyon.techntrek.in/learn/backpropagation"},{"slug":"optimizers","title":"Optimisers: SGD, momentum and Adam","url":"https://tachyon.techntrek.in/learn/optimizers"}],"read":[],"sources":[{"title":"Understanding the difficulty of training deep feedforward neural networks","url":"https://proceedings.mlr.press/v9/glorot10a.html","org":"Glorot and Bengio, AISTATS","date":"2010"},{"title":"Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification","url":"https://arxiv.org/abs/1502.01852","org":"He et al.","date":"2015"}],"tags":["neural networks","training","interview"]},{"slug":"agentic-architectures-poster","url":"https://tachyon.techntrek.in/trending/2026-09-11#agentic-architectures-poster","date":"2026-09-11","kind":"concept","title":"3 agent patterns on one graphic: CodeAct, ReAct and agentic RAG. Two of the three descriptions oversell.","said":"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).","takeaway":"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.","forBusiness":"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.","numbers":[],"verdicts":{"holds":2,"partly":1,"wrong":0,"unverified":0,"opinion":0},"checks":[{"claim":"CodeAct consolidates an agent's actions into executable Python code.","verdict":"holds","note":"That is the paper's proposal: one action space, executable code, which lets the agent compose tools and loop instead of emitting one JSON call at a time. It reported up to a 20 percent higher success rate than JSON or text actions on their benchmark.","source":{"title":"Executable Code Actions Elicit Better LLM Agents","url":"https://arxiv.org/abs/2402.01030"}},{"claim":"ReAct overcomes hallucination and error propagation in chain-of-thought.","verdict":"partly","note":"The ReAct paper shows fewer hallucinated facts on HotpotQA and FEVER because reasoning can consult a Wikipedia tool, and it reports the remaining failure modes honestly. It reduces the problem; it does not remove it.","source":{"title":"ReAct: Synergizing Reasoning and Acting in Language Models","url":"https://arxiv.org/abs/2210.03629"}},{"claim":"Agentic RAG adds agents into the RAG pipeline to orchestrate its components.","verdict":"holds","note":"A fair description of the pattern: the model decides when to retrieve, from where, whether to re-query and when to stop, instead of a fixed retrieve-then-answer step."}],"learn":[{"slug":"agent-loop","title":"The agent loop","url":"https://tachyon.techntrek.in/learn/agent-loop"},{"slug":"tool-calling","title":"Tool calling","url":"https://tachyon.techntrek.in/learn/tool-calling"},{"slug":"rag","title":"Retrieval-augmented generation (RAG)","url":"https://tachyon.techntrek.in/learn/rag"},{"slug":"agent-memory","title":"Agent memory","url":"https://tachyon.techntrek.in/learn/agent-memory"}],"read":[{"label":"Single agent with tools","url":"https://tachyon.techntrek.in/architectures/single-agent-tools"},{"label":"Agentic graph RAG","url":"https://tachyon.techntrek.in/architectures/agentic-graph-rag"}],"sources":[{"title":"ReAct: Synergizing Reasoning and Acting in Language Models","url":"https://arxiv.org/abs/2210.03629","org":"Yao et al., Princeton and Google","date":"2022"},{"title":"Executable Code Actions Elicit Better LLM Agents","url":"https://arxiv.org/abs/2402.01030","org":"Wang et al., UIUC","date":"2024"}],"tags":["agents","RAG","architecture"]},{"slug":"amazon-multi-agent-and-stac","url":"https://tachyon.techntrek.in/trending/2026-09-11#amazon-multi-agent-and-stac","date":"2026-09-11","kind":"research","title":"91.2 percent: how often innocent tool calls could be chained into a harmful action in Amazon's STAC study.","said":"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).","takeaway":"Agreement between agents is not accuracy, and per-tool safety is not chain safety. Evaluate the sequence, not the pieces.","forBusiness":"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.","numbers":[{"value":"91.2%","label":"mean final attack success (STAC)"},{"value":"483","label":"generated attack chains"}],"verdicts":{"holds":2,"partly":0,"wrong":0,"unverified":0,"opinion":0},"checks":[{"claim":"Injecting social relation priors (attitude, trust, influence) into multi-agent systems drives agreement without reliably improving correctness.","verdict":"holds","note":"The paper (arXiv, 4 August 2026) finds relational positivity acts as convergence pressure: agents agree more, and in objective question answering that agreement can rise while correctness-conditioned agreement does not. The authors recommend comparing against a no-prior baseline before using such priors.","source":{"title":"Relational Priors as Convergence Pressure in LLM-Based Multi-Agent Systems","url":"https://arxiv.org/abs/2608.03239"}},{"claim":"Individually benign tools can be composed into chains that let an agent do harm no single tool could.","verdict":"holds","note":"STAC (Sequential Tool Attack Chaining) auto-generated 483 attack cases across 1,352 interactions; the mean final attack success rate was 91.2 percent while the attacks stayed almost undetectable to the agents' own checks.","source":{"title":"STAC: When Innocent Tools Form Dangerous Chains for LLM Agents","url":"https://arxiv.org/abs/2509.25624"}}],"learn":[{"slug":"prompt-injection","title":"Prompt injection","url":"https://tachyon.techntrek.in/learn/prompt-injection"},{"slug":"guardrails","title":"Guardrails","url":"https://tachyon.techntrek.in/learn/guardrails"},{"slug":"agent-loop","title":"The agent loop","url":"https://tachyon.techntrek.in/learn/agent-loop"},{"slug":"mcp","title":"Model Context Protocol (MCP)","url":"https://tachyon.techntrek.in/learn/mcp"}],"read":[],"sources":[{"title":"Relational Priors as Convergence Pressure in LLM-Based Multi-Agent Systems","url":"https://arxiv.org/abs/2608.03239","org":"Shen et al., Amazon","date":"2026"},{"title":"STAC: When Innocent Tools Form Dangerous Chains for LLM Agents","url":"https://arxiv.org/abs/2509.25624","org":"Li et al., Amazon","date":"2025"}],"tags":["agents","safety","multi-agent","research"]},{"slug":"how-to-build-an-ai-agent-aug-2026","url":"https://tachyon.techntrek.in/trending/2026-09-11#how-to-build-an-ai-agent-aug-2026","date":"2026-09-11","kind":"tools","title":"8 steps and 12 products on one graphic. The model names check out; the integration counts do not.","said":"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.","takeaway":"The stack in the table is real. The step that decides the outcome, evals, is listed last and should come first.","forBusiness":"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.","numbers":[],"verdicts":{"holds":2,"partly":0,"wrong":0,"unverified":1,"opinion":1},"checks":[{"claim":"ChatGPT runs GPT-5.6.","verdict":"holds","note":"OpenAI released GPT-5.6 on 9 July 2026 in three variants (Sol, Terra, Luna) across ChatGPT, Codex and the API.","source":{"title":"GPT-5.6: Frontier intelligence that scales with your ambition","url":"https://openai.com/index/gpt-5-6/"}},{"claim":"Windsurf runs SWE-1.6.","verdict":"holds","note":"Cognition shipped SWE-1.6 in April 2026, served at up to 950 tokens per second on Cerebras for paying users.","source":{"title":"Introducing SWE 1.6: Improving Model UX","url":"https://cognition.com/blog/swe-1-6"}},{"claim":"Lindy has 1,000+ integrations, Zapier 9,000+ apps, n8n 500+.","verdict":"unverified","note":"Vendor marketing counts change monthly and we did not confirm them against the vendors' current pages. Treat them as order-of-magnitude."},{"claim":"The eight-step order is the right order.","verdict":"opinion","note":"It is a sensible order with one weakness: testing and evals sit last. In practice the eval set is written at step one, from the success criteria, or you cannot tell whether steps two to seven worked."}],"learn":[{"slug":"agent-loop","title":"The agent loop","url":"https://tachyon.techntrek.in/learn/agent-loop"},{"slug":"agent-memory","title":"Agent memory","url":"https://tachyon.techntrek.in/learn/agent-memory"},{"slug":"tool-calling","title":"Tool calling","url":"https://tachyon.techntrek.in/learn/tool-calling"},{"slug":"mcp","title":"Model Context Protocol (MCP)","url":"https://tachyon.techntrek.in/learn/mcp"},{"slug":"evals","title":"Evals","url":"https://tachyon.techntrek.in/learn/evals"}],"read":[{"label":"How we work: the six stages","url":"https://tachyon.techntrek.in/process"}],"sources":[{"title":"GPT-5.6: Frontier intelligence that scales with your ambition","url":"https://openai.com/index/gpt-5-6/","org":"OpenAI","date":"2026-07-09"},{"title":"Introducing SWE 1.6: Improving Model UX","url":"https://cognition.com/blog/swe-1-6","org":"Cognition","date":"2026"}],"tags":["agents","tools","models"]},{"slug":"navier-stokes-and-unstable-singularities","url":"https://tachyon.techntrek.in/trending/2026-09-11#navier-stokes-and-unstable-singularities","date":"2026-09-11","kind":"viral","title":"The Navier–Stokes equations are not 'one of the hardest ever solved'. They are unsolved, and AI just found new singularities.","said":"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.","takeaway":"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.","forBusiness":"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.","numbers":[],"verdicts":{"holds":1,"partly":1,"wrong":1,"unverified":0,"opinion":0},"checks":[{"claim":"C. Navier, G. Stokes, 1845.","verdict":"partly","note":"Navier published the viscous equations in 1822; Stokes gave the modern derivation in 1845. One date for both names hides 23 years."},{"claim":"Still one of the hardest equations ever solved.","verdict":"wrong","note":"Engineers solve them numerically every day, but whether smooth solutions always exist in three dimensions is an open Clay Millennium Prize problem, unsolved since it was posed in 2000. Unsolved is the honest word.","source":{"title":"Navier–Stokes Equation (Millennium Prize problem)","url":"https://www.claymath.org/millennium/navier-stokes-equation/"}},{"claim":"AI is now finding fluid singularities (the vortex post).","verdict":"holds","note":"In September 2025 Google DeepMind with NYU, Stanford and Brown reported the first systematic discovery of unstable singularities in fluid equations (including 3D Euler with boundary), using physics-informed neural networks and a high-precision Gauss–Newton optimiser to near machine precision. It is a step towards the Navier–Stokes question, not an answer to it.","source":{"title":"Discovery of Unstable Singularities","url":"https://arxiv.org/abs/2509.14185"}}],"learn":[{"slug":"neural-networks","title":"Neural networks","url":"https://tachyon.techntrek.in/learn/neural-networks"},{"slug":"loss-and-gradient-descent","title":"Loss and gradient descent","url":"https://tachyon.techntrek.in/learn/loss-and-gradient-descent"},{"slug":"optimizers","title":"Optimisers: SGD, momentum and Adam","url":"https://tachyon.techntrek.in/learn/optimizers"}],"read":[],"sources":[{"title":"Navier–Stokes Equation","url":"https://www.claymath.org/millennium/navier-stokes-equation/","org":"Clay Mathematics Institute","date":"2000"},{"title":"Discovery of Unstable Singularities","url":"https://arxiv.org/abs/2509.14185","org":"Wang et al., Google DeepMind and universities","date":"2025-09-17"},{"title":"Discovering new solutions to century-old problems in fluid dynamics","url":"https://deepmind.google/blog/discovering-new-solutions-to-century-old-problems-in-fluid-dynamics/","org":"Google DeepMind","date":"2025"}],"tags":["AI for science","maths","viral"]},{"slug":"resnext-cardinality","url":"https://tachyon.techntrek.in/trending/2026-09-11#resnext-cardinality","date":"2026-09-11","kind":"concept","title":"32 parallel paths beat one wider block: ResNeXt's 'cardinality', doing the rounds again in 2026.","said":"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.","takeaway":"Split, transform, merge. The same idea now runs in mixture-of-experts language models: many small parallel paths instead of one big one.","forBusiness":"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.","numbers":[],"verdicts":{"holds":2,"partly":0,"wrong":0,"unverified":0,"opinion":0},"checks":[{"claim":"Cardinality (the size of the set of transformations) is an essential dimension alongside depth and width.","verdict":"holds","note":"The paper shows on ImageNet-1K that raising cardinality improves accuracy at the same complexity, and does so more effectively than going deeper or wider.","source":{"title":"Aggregated Residual Transformations for Deep Neural Networks","url":"https://arxiv.org/abs/1611.05431"}},{"claim":"ResNeXt was the foundation of the ILSVRC 2016 entry that took second place.","verdict":"holds","note":"Stated in the abstract; the models and code were released publicly.","source":{"title":"Aggregated Residual Transformations for Deep Neural Networks","url":"https://arxiv.org/abs/1611.05431"}}],"learn":[{"slug":"convolution-and-vision","title":"Convolution and vision","url":"https://tachyon.techntrek.in/learn/convolution-and-vision"},{"slug":"residual-stream-and-normalization","title":"Residual stream and normalisation","url":"https://tachyon.techntrek.in/learn/residual-stream-and-normalization"},{"slug":"mixture-of-experts","title":"Mixture of experts","url":"https://tachyon.techntrek.in/learn/mixture-of-experts"}],"read":[],"sources":[{"title":"Aggregated Residual Transformations for Deep Neural Networks","url":"https://arxiv.org/abs/1611.05431","org":"Xie et al., UC San Diego and Facebook AI Research, CVPR","date":"2017"}],"tags":["vision","architecture","papers"]},{"slug":"adobe-foundation-model-data-hiring","url":"https://tachyon.techntrek.in/trending/2026-09-11#adobe-foundation-model-data-hiring","date":"2026-09-11","kind":"hiring","title":"Billions of images and videos a month: Adobe is hiring for the data pipeline behind Firefly, not the model.","said":"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.","takeaway":"Frontier labs spend as much engineering on inference for data curation as on training. The unglamorous pipeline is where the senior roles are.","forBusiness":"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.","numbers":[],"verdicts":{"holds":0,"partly":0,"wrong":0,"unverified":0,"opinion":1},"checks":[{"claim":"The bottleneck for a foundation model team is inference throughput on the data pipeline.","verdict":"opinion","note":"A hiring signal rather than a claim to check: captioning, filtering and scoring billions of items with vision-language models is itself a giant inference workload, so the pipeline job is an inference optimisation job."}],"learn":[{"slug":"diffusion-models","title":"Diffusion models","url":"https://tachyon.techntrek.in/learn/diffusion-models"},{"slug":"prefill-and-decode","title":"Prefill and decode","url":"https://tachyon.techntrek.in/learn/prefill-and-decode"},{"slug":"speculative-decoding","title":"Speculative decoding","url":"https://tachyon.techntrek.in/learn/speculative-decoding"},{"slug":"batching-and-paged-attention","title":"Batching and PagedAttention","url":"https://tachyon.techntrek.in/learn/batching-and-paged-attention"}],"read":[{"label":"Visual document retrieval architecture","url":"https://tachyon.techntrek.in/architectures/visual-document-retrieval"}],"sources":[],"tags":["hiring","vision","data pipelines","Adobe"]},{"slug":"mistral-pretraining-hiring","url":"https://tachyon.techntrek.in/trending/2026-09-11#mistral-pretraining-hiring","date":"2026-09-11","kind":"hiring","title":"Thousands of GPUs, petabytes per run, no ML experience required: Mistral's pretraining team wants systems engineers.","said":"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.","takeaway":"The model is a few thousand lines; the training system is the product. Reliability and throughput engineers are the scarce hires.","forBusiness":"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.","numbers":[],"verdicts":{"holds":0,"partly":0,"wrong":0,"unverified":0,"opinion":1},"checks":[{"claim":"Frontier pretraining is mostly a distributed-systems problem.","verdict":"opinion","note":"The post says so explicitly, and it matches how the large labs staff these teams: the recipe is small, the infrastructure that keeps 10,000 GPUs busy through hardware failures is not."}],"learn":[{"slug":"pretraining-and-scaling-laws","title":"Pretraining and scaling laws","url":"https://tachyon.techntrek.in/learn/pretraining-and-scaling-laws"},{"slug":"distillation","title":"Distillation","url":"https://tachyon.techntrek.in/learn/distillation"},{"slug":"fine-tuning-and-lora","title":"Fine-tuning and LoRA","url":"https://tachyon.techntrek.in/learn/fine-tuning-and-lora"}],"read":[],"sources":[],"tags":["hiring","pretraining","infrastructure","Mistral"]},{"slug":"ai-engineer-job-description","url":"https://tachyon.techntrek.in/trending/2026-09-11#ai-engineer-job-description","date":"2026-09-11","kind":"hiring","title":"One job post, five nouns: data pipelines, RAGs, MCPs, evals and human approval steps.","said":"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.","takeaway":"The market's definition of an AI engineer in 2026: someone who can build the workflow and prove it works, then hand it over.","forBusiness":"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.","numbers":[],"verdicts":{"holds":0,"partly":0,"wrong":0,"unverified":0,"opinion":1},"checks":[{"claim":"This is what buyers now ask for.","verdict":"opinion","note":"A hiring signal. The notable part is the order: evals, approval steps, monitoring and documentation are listed as core duties, not extras, and the application asks how you evaluated your project."}],"learn":[{"slug":"rag","title":"Retrieval-augmented generation (RAG)","url":"https://tachyon.techntrek.in/learn/rag"},{"slug":"mcp","title":"Model Context Protocol (MCP)","url":"https://tachyon.techntrek.in/learn/mcp"},{"slug":"evals","title":"Evals","url":"https://tachyon.techntrek.in/learn/evals"},{"slug":"guardrails","title":"Guardrails","url":"https://tachyon.techntrek.in/learn/guardrails"},{"slug":"llm-observability","title":"LLM observability","url":"https://tachyon.techntrek.in/learn/llm-observability"}],"read":[{"label":"How we work","url":"https://tachyon.techntrek.in/process"},{"label":"No lock-in, full documentation","url":"https://tachyon.techntrek.in/#no-lock-in"}],"sources":[],"tags":["hiring","evals","RAG","MCP"]},{"slug":"system-design-canon-40-problems","url":"https://tachyon.techntrek.in/trending/2026-09-11#system-design-canon-40-problems","date":"2026-09-11","kind":"roadmap","title":"40 system design problems in 8 groups, from 'design Twitter' to the boss fights. Now a track on our roadmap.","said":"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).","takeaway":"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.","forBusiness":"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.","numbers":[],"verdicts":{"holds":0,"partly":0,"wrong":0,"unverified":0,"opinion":1},"checks":[{"claim":"These 40 are the canon.","verdict":"opinion","note":"They overlap almost entirely with the problem sets in the standard interview books and courses, so as a study list it is representative. None of them has a single right answer, which is the point."}],"learn":[{"slug":"hnsw","title":"HNSW vector search","url":"https://tachyon.techntrek.in/learn/hnsw"},{"slug":"bm25","title":"BM25 keyword search","url":"https://tachyon.techntrek.in/learn/bm25"},{"slug":"hybrid-search-and-reranking","title":"Hybrid search and reranking","url":"https://tachyon.techntrek.in/learn/hybrid-search-and-reranking"},{"slug":"kv-cache","title":"KV cache","url":"https://tachyon.techntrek.in/learn/kv-cache"}],"read":[{"label":"System design track on the roadmap","url":"https://tachyon.techntrek.in/learn/roadmap?track=system-design"},{"label":"Data engineering and cloud","url":"https://tachyon.techntrek.in/services/data-engineering-cloud"}],"sources":[],"tags":["system design","roadmap","interview"]}]}]}