๐ฅ Top Story
Grok Build vs. Claude Code: I tested which one has the better memory
On September 16, xAI added memory to Grok Build, its terminal coding agent, storing markdown notes on conventions and decisions at both a per-project (workspace) scope and a global scope that applies across all projects. Claude Code has offered similar "auto memory" for months, keeping a MEMORY.md index plus per-note files scoped to each repository. In a head-to-head test, Grok Build's global scope correctly carried a "for all my projects" rule into an unrelated second repo, while Claude Code's per-repo scoping failed to apply it despite warning the user. Claude Code was faster in every test, but used more total tokens (576,863 vs. 390,848) and cost roughly 2-3x more per session ($1.05 vs. $0.41). The author concluded Grok Build is the better choice for most people right now on accuracy and cost.
๐ก Why it matters: For teams standardizing on a terminal coding agent, memory scope design (global vs. per-repo) and token cost per session matter as much as raw speed when deciding which tool to give engineers.
๐ Read more ยท The New Stack
Kubernetes & Cloud Native
How Ramp runs GPU AI workloads at scale with ECS Managed Instances
AWS Containers
Ramp, the fintech platform, runs GPU-powered AI inference continuously. It migrated its GPU workloads from self-managed EC2 onto Amazon ECS Managed Instances, moving roughly 50 to 60 EC2 instances with a Terraform-based implementation, according to the article's own excerpt. A related AWS customer case study describing the same migration states Ramp cut task provisioning times by more than 50%. This sped up deployments while offloading infrastructure management to AWS. The source blog post itself could not be fetched due to a network restriction, so this summary relies on the given excerpt plus that separate AWS case-study page rather than the full article text.
๐ก Even without confirming every detail from the original post, the pattern is a concrete signal that ECS Managed Instances is becoming viable for GPU inference at production scale, letting teams keep Terraform-driven EC2 flexibility while shedding node-provisioning toil.
Kubernetes v1.37: Tracking When a PersistentVolumeClaim Was Last Used (Beta)
Kubernetes
Kubernetes v1.37 promotes the PersistentVolumeClaimUnusedSinceTime feature gate from Alpha (v1.36) to Beta, enabled by default. The existing PVC protection controller now adds an "Unused" condition to each PVC's status. It flips to True (Reason=NoPodsUsingPVC) when no non-terminal pod references the claim, and back to False (Reason=PodUsingPVC) once a pod uses it again. The condition's standard lastTransitionTime field records the exact moment a PVC became idle, so cluster admins, monitoring tools, or external controllers can query it directly instead of writing custom scripts that cross-reference PVCs and pods. Terminated pods don't count as usage, but pending or even unschedulable pods do; Kubernetes itself doesn't delete anything based on this -- it just exposes the idle timestamp for cleanup tooling to act on.
๐ก With a native Unused condition and timestamp, teams can now build storage cost-governance and stale-PVC cleanup automation directly against the Kubernetes API instead of maintaining bespoke PVC-to-pod cross-referencing scripts, cutting both engineering effort and the risk of false-positive deletions from custom heuristics.
AI & ML
Pruning LLMs Like a Physicist: Block Removal as an Ising Optimization Problem
Hugging Face
Multiverse Computing frames LLM block (layer) removal as finding the low-energy state of an Ising spin-glass system, where each transformer block is a binary keep/drop variable. Pairwise couplings between blocks are computed from a Hessian via a second-order Taylor expansion built from a single calibration pass, and brute-force, tabu-search, or quantum-inspired solvers then search for the best block combination. Unlike mean-field heuristics that score blocks independently, this captures block-to-block interactions, and the authors note the best-performing pruning configuration is often an excited state of the spin system rather than the ground state. At 50% depth compression of Llama-3.3-70B-Instruct, the method reportedly holds MMLU around 77 versus the mid-50s for the strongest baseline โ nearly a 23-point gain โ and it generalizes to hybrid Mamba/attention/MoE architectures like NVIDIA's Nemotron family. Note: the huggingface.co source page could not be directly fetched due to a network restriction; this summary is based on the underlying paper and press coverage found via search.
๐ก If a pruning method can hold near-baseline MMLU at 50% depth compression, it directly shrinks both inference cost and the serving fleet size needed for a given model, so it's worth adding to the model-selection checklist before deployment rather than treated purely as a research curiosity.
Advisory Group on Mathematics and Artificial Intelligence
OpenAI
OpenAI announced it is working with an independent Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study in Princeton. The nine-member group consists of mathematicians Francois Charles, Camillo De Lellis, Timothy Gowers, Martin Hairer, Nikhil Srivastava, Ulrike Tillmann, Ravi Vakil, Edward Witten, and Melanie Matchett Wood. They serve unpaid, can offer unsolicited advice, and can go public with their views, giving them a degree of independence from the company. The group's mandate is to advise on how OpenAI assesses and communicates emerging mathematical results, upholds academic and professional research standards, and builds tools that support mathematical research and learning, but it is explicitly not empowered to advise on or slow OpenAI's pace of internal mathematics research. The announcement follows claims that an internal OpenAI model resolved more than 100 long-standing open problems across most areas of mathematics, which drew concern from parts of the mathematical community.
๐ก This is a governance layer where outside experts review how AI-generated mathematical claims get vetted and communicated, which is a useful signal for ops teams to treat any vendor's AI performance or research claims as needing independent validation before they factor into adoption decisions.
Higgsfield AI ships new video features in a day with GPT-6 Astra
OpenAI
Higgsfield AI used OpenAI's GPT-6 Astra to ship a new video ad feature in a single day. GPT-6 Astra is described as OpenAI's strongest model yet for coding, reasoning, and complex knowledge work. It is paired with real-time 3D generation and long-horizon task planning that lets it plan work across multiple steps. Higgsfield built a feature on top of it that generates up to 100 ad variations from a single prompt, including versions tailored to different countries. A single engineer delivered the entire release in one day, letting small businesses get video ads to market faster.
๐ก A single engineer shipping a production feature in a day by leaning on a model's long-horizon task planning suggests that deployment pipelines and review processes sized around traditional team throughput may need to be re-evaluated once AI coding agents are in the loop.
Building standards for the next phase of AI
OpenAI
In a post published September 21, OpenAI proposed a US-led international effort to build common technical standards for frontier AI. The plan leans on the existing network of AI safety institutes and national bodies like CAISI to manage the benefits and risks of automated AI research, including recursive self-improvement (RSI). Concretely, it calls for shared measurement and incident-reporting protocols: tracking self-improvement progress, the share of autonomous research happening inside AI labs, thresholds that trigger human review, and a common system for classifying and reporting alignment incidents. OpenAI stressed these would be voluntary technical benchmarks, not licenses or mandatory pre-release review, leaving it to each government whether to write them into law. The company stated the US "must lead efforts to establish technological standards for cutting-edge AI in cooperation with countries around the world." The source page could not be fetched directly (network block); this is written from verified excerpts of the original article found via search.
๐ก If self-improvement thresholds and incident-reporting formats become standardized, internal AI platform teams should expect to need human-review gates and incident-classification logging built into their model deployment pipelines.
tokenizers v1: encode, decode and scaling, measured
Hugging Face
Hugging Face released tokenizers v1, a major new version of its tokenizer library, on September 21, with the release centered on performance. The v1 release candidate is described as being faster than the previous v0.23 in some cases, by "tens of times." The blog post presents per-area benchmark results covering encoding, decoding, scaling, latency, throughput, memory, and package size. The underlying motivation is that tokenization, though computationally light relative to model inference, can become a bottleneck that starves the model of data as workloads scale. This includes training on massive datasets, serving concurrent requests, or processing long inputs. The source page could not be fetched directly (network block); this is written from verified excerpts of the original article found via search.
๐ก In high-throughput or long-context serving setups, simply upgrading the tokenizer version could reduce GPU idle time caused by data starvation, so it's worth tracking tokenizer version and performance as part of inference-pipeline observability.
Cloud Updates
Global AI routing with <1% overhead on multi-cluster GKE Inference Gateway
Google Cloud
Google Cloud says its multi-cluster GKE Inference Gateway keeps routing overhead under 1% while distributing AI traffic across geographically scattered GPU/TPU clusters. Instead of network-layer round-robin, the LLM-d router and an Endpoint Picker Proxy (EPP) read live KV-cache utilization from each inference engine. Once that crosses a 40% threshold, traffic spills over to the next healthy region. In a test spanning three GKE clusters (us-east5, us-west8, europe-west4) with 17,000 nodes running an MoE model served via SGLang, throughput scaled near-linearly from 0.72 req/s (2,898 tok/s) on one cluster to 2.10 req/s (8,457 tok/s) on three, with success rates around 99.9%. The system also integrates with Kubernetes LeaderWorkerSet (LWS) to respect master-worker pod topologies used by distributed LLM serving engines.
๐ก As accelerator scarcity becomes the norm, treating application-level signals like KV-cache utilization as the routing input โ rather than relying on single-region capacity โ is what actually turns fragmented multi-cluster GPU/TPU capacity into usable throughput.
Maximizing Apache Spark availability: Mitigating compute stockouts with flexible VMs and other best practices
Google Cloud
Google Cloud describes 'Flexible VMs' in Managed Service for Apache Spark as a way to survive machine-type capacity stockouts caused by surging AI demand. Instead of pinning a cluster to one machine family, operators define a ranked list mixing Gen2 (N2, N2D) and Gen4 (N4, C4) families, and Managed Spark automatically tries provisioning in rank order, falling through to the next tier on failure. A sample configuration ranks n2d-standard-16/n2-standard-16 first, falls back to n4-standard-16/n4d-standard-16 on Hyperdisk Balanced, then c4/c3-standard machines, and finally e2-standard-16 as a last resort, all set via the --worker-instance-selection flag on gcloud dataproc clusters create. The rule applies to primary workers, preemptible secondary workers, and master nodes alike. Companion practices include AutoZone, autoscaling, partial cluster creation, regional fallback, and adopting machine-family-agnostic flexible Committed Use Discounts instead of family-specific CUDs.
๐ก Any Spark pipeline hard-pinned to a single machine family turns a regional stockout into an SLA breach, so baking rank-ordered machine fallbacks and family-agnostic CUDs into cluster definitions now is the difference between a graceful degrade and a failed run later.
Scale your AI workloads faster and more efficiently with GKE Pod snapshots
Google Cloud
GKE Pod snapshots let a workload's fully loaded state, including CPU and GPU memory, be saved and restored on demand, eliminating the cold-start initialization phase entirely. Snapshots persist to high-throughput Cloud Storage, and a Pod snapshot CRD declaratively defines which pods to snapshot, the storage location, retention policy, and restore behavior. Google's benchmarks cite an 89% reduction in AI inference startup latency. A 70-billion-parameter LLM loads in 37 seconds and an 8-billion-parameter model in 15 seconds on A3 H100 GPUs. As a real-world example, customer Retake (Codeway) reportedly cut startup time from one minute down to 8 seconds using the feature.
๐ก Because per-replica model loading currently scales cost and latency linearly with fleet size, wiring Pod snapshots into autoscaling policy is what lets teams handle traffic spikes with fast restores instead of paying to keep idle over-provisioned capacity warm.
Python Workers are now generally available
Cloudflare
Cloudflare has made Python Workers generally available, putting Python on equal first-class footing with TypeScript/JavaScript on its Developer Platform. FastAPI, Django, and Flask are natively supported via the workers.asgi and workers.wsgi connector packages, with the Workers runtime itself acting as the web server so no Uvicorn or Gunicorn setup is needed. Because HTTP clients were patched to route through JavaScript's fetch API, AI libraries such as openai, langchain, and mcp work natively without glue code, and Python Workers integrate directly with Workers AI, R2, D1, Hyperdrive, Durable Objects, Queues, and Workflows. Cloudflare also proposed and got accepted PEP 783 (PyEmscripten), standardizing cross-compilation of Python packages to WebAssembly, and added corresponding support to cibuildwheel. Note: the blog.cloudflare.com source page could not be directly fetched due to a network restriction; this summary is based on related coverage found via search.
๐ก Running FastAPI, Django, or Flask on the edge without standing up a separate WSGI/ASGI server means existing Python backend teams can shift their deployment footprint โ region count, cold starts, ops overhead โ to edge compute without a rewrite, which is reason enough to move edge migration up the roadmap.
Turning security complexity into useful intelligence: Whatโs new in Red Hat Lightspeed
Red Hat
Red Hat introduced a new security capability in Lightspeed (formerly Red Hat Insights). The core change: instead of surfacing raw YARA signature output, a detection now comes with an immediate, plain-language summary explaining exactly what the signature detects and why it's a risk for that specific workload. The motivation is the cybersecurity skills gap โ infrastructure generalists are effectively expected to act as security experts, spending time decoding complex detection data from standard alerts before they can even decide on next steps. Lightspeed's broader goal is to use predictive analytics and AI-driven insights so teams can operate faster and more securely at scale. The source page could not be fetched directly (network block); this is written from verified excerpts of the original article found via search.
๐ก Translating raw detection signatures into workload-specific plain-language context can meaningfully cut mean-time-to-respond for infra teams that lack dedicated security staff.
2026 update: The road to quantum-safe cryptography in Red Hat OpenShift
Red Hat
Red Hat published a 2026 update to a post-quantum cryptography roadmap for OpenShift it first wrote a year earlier. Because OpenShift only builds on RHEL's EUS releases, quantum-safe support in core OpenShift components has to wait for the RHEL 9.8 and 10.2 releases planned for spring 2026. In the meantime, UBI 9.7, UBI 10, and UBI 10.1 already ship PQC-capable OpenSSL supporting ML-KEM, ML-DSA, and SLH-DSA, letting layered applications and workloads built on those UBI images start adopting post-quantum cryptography sooner. Red Hat also noted it explored bringing quantum-safe elements to OpenShift in 2025 via the ingress controller, Service Mesh, or frontend proxies using quantum-safe TLS key exchange. A cited obstacle is that many OpenShift components are written in Go, which is tightly coupled to specific Kubernetes upstream and OpenShift version pairings. The source page could not be fetched directly (network block); this is written from verified excerpts of the original article found via search.
๐ก Since core-component PQC support is gated on RHEL 9.8/10.2 (spring 2026), cluster teams can get a head start on quantum-safe posture now by migrating workloads onto the PQC-capable UBI images rather than waiting for the platform-wide upgrade.
DevOps & Infrastructure
TypeSafe launched Jev because sequential LLMs are โtotally useless for computersโ
The New Stack
TypeSafe emerged from two years of stealth last week with $40 million in seed funding led by DCVC. It launched Jev, a "System One" model built for machine decisions rather than chat. Founder Almeida argues that autoregressive, sequential text generation is "great for a natural conversation, but totally useless for computers," and says the idea came from asking why superhuman chat models haven't produced AGI. Jev outputs decisions as probabilities and confidence scores instead of words. TypeSafe claims this makes it immune to hallucination because it behaves "a lot more like code." The company's launch post claims Jev is 20-200x faster and 40-400x cheaper than conventional LLMs, with output tokens free.
๐ก If TypeSafe's decision-model claims hold up under independent benchmarking, routing high-volume, low-latency agent decisions to a specialized non-chat model instead of a general LLM could meaningfully cut inference cost and latency in production pipelines, but the "can't hallucinate" claim from a two-year-stealth startup needs third-party validation before it's trusted for critical decisions.
Your AI agent is burning tokens on choices that donโt need words
The New Stack
Kev is a new open-source family of decision models -- 0.8B, 4B, and 9B parameters, built on Qwen 3.5. It skips text generation entirely for simple agent decisions like tool routing, yes/no checks, and ranking. Instead of an autoregressive decoding loop, Kev processes the state and candidates in a single forward pass and reads probabilities off a pointer head. It supports three decision types -- Noul (yes/no), Choice (pick among candidates), and Score (ordered levels) -- mirroring TypeSafe's System One API. Kev-9B reached 83.7% accuracy on an out-of-domain test, though the authors note probability calibration can drift on unseen data and fine-tuning reduces the base model's general-knowledge and arithmetic performance.
๐ก Swapping small non-generative decision models in for full LLM calls at discrete agent decision points (routing, gating, ranking) is a concrete lever for cutting per-step token spend in multi-agent pipelines, but the calibration drift on unseen data means it needs monitoring, not blind trust, at each new decision domain.
Open-Sourcing Rebalancer: A Generic, High-Performance Library for Solving Assignment Problems
Meta Engineering
Meta open-sourced Rebalancer on September 21, 2026 under the Apache 2.0 license. It is a C++ library (with Python bindings) for solving assignment problems -- allocating objects to containers under hard constraints while optimizing an objective -- that has run in production at Meta for more than nine years. The library separates four concerns: how to specify a problem via a named-spec DSL (objects, containers, dimensions, goals, constraints), how to store it efficiently in memory, how to solve it, and how to debug it, and it offers both a local-search engine for large-scale instances (up to roughly 1 million objects/containers) and an optimal MIP backend using HiGHS, Gurobi, or FICO Xpress. Inside Meta it has been used for hardware and server allocation, ML training/inference placement, traffic routing, and load-balancing migrations, as well as non-infrastructure cases like assigning meeting rooms or support tickets. The design and production experience behind it were documented in a 2024 OSDI paper, "Optimizing Resource Allocation in Hyperscale Datacenters."
๐ก For infra teams building capacity placement, load-balancing, or traffic-routing logic in-house, Rebalancer offers a production-hardened, Apache-2.0-licensed alternative to hand-rolled bin-packing heuristics, with a local-search path for scale and a MIP path for provably optimal results when problem size allows.
Introducing Our New Dropbox API Documentation
Dropbox
Dropbox has launched a redesigned API documentation site at docs.dropboxapi.com. It brings endpoints, types, schemas, and related docs together in one responsive interface with improved navigation and search. It also adds an in-browser "try it" feature that lets developers call endpoints and see real API responses without leaving the page. It also ships an embedded AI assistant that can answer questions about endpoints, help troubleshoot, and provide code examples in place, plus support for connecting compatible AI tools via an MCP (Model Context Protocol) server. Note: the source page itself could not be directly fetched due to a network restriction; this summary is based on publicly available search results describing the announcement.
๐ก Live, testable docs plus MCP integration make it easy to wire the Dropbox API directly into coding assistants and agent workflows, so teams adopting it should audit credential scope and request logging up front before agents start calling the API on their behalf.
How Adaptive Tail Sampling Works in the OpenTelemetry Collector
Honeycomb
Honeycomb has donated its adaptive tail sampling processor, built on years of running Refinery, to the OpenTelemetry Collector as adaptivetailsamplingprocessor in opentelemetry-collector-contrib (currently working toward alpha); it's already usable today via the Honeycomb Collector Distribution, a drop-in replacement for the contrib image. The processor buffers whole traces instead of deciding per-span, firing its decision decision_delay (2 seconds by default) after the root span arrives so late-arriving spans still get counted, with trace_timeout (30 seconds by default) as a safety net that forces a decision for traces that never see a root span. Decisions are routed via first-match rules to adaptive samplers such as an EMA (Exponential Moving Average) sampler, which recalculates sample rates per adjustment interval based on traffic frequency so it can hold onto important context during spikes without blowing budgets. Compared with Refinery, this processor brings the same tail-sampling philosophy into the standard Collector pipeline, letting teams approximate Refinery-style dynamic sampling without standing up separate Refinery infrastructure. Note: WebFetch could not open the source page directly, so this summary is built only from information confirmed through search-engine snippets of the article and related Honeycomb posts (parameter defaults, processor architecture, Refinery comparison).
๐ก Running trace-level adaptive sampling inside the standard Collector is attractive, but since it is still alpha and buffers entire traces, capacity planning needs to account for the memory and ingestion-latency cost that scales with the decision_delay and trace_timeout settings.
Inside Petal: Building the Worldโs First Petabit-Class Transoceanic Subsea Cable
Meta Engineering
Meta announced Petal, a subsea cable connecting France and the United States over roughly 7,000 km, which it says will be the first subsea cable system to deliver petabit-class capacity at transoceanic distances. Petal will carry 1 petabit per second, doubling the capacity of today's best-in-class transoceanic cables without a proportional increase in power or physical infrastructure. This marks the largest generational capacity jump for any transoceanic subsea system. It is the first system to deploy multi-core fiber at transoceanic distance, using two-core fiber with two optical transmission paths inside a single fiber alongside an advanced single-housing repeater design that cuts power consumption. Meta is partnering with subsea cable specialists NEC and Sumitomo Electric Industries to build it, Orange will support the landing on France's Atlantic coast, and the cable is expected to enter service in 2029.
๐ก With a 2029 service date this has no near-term capacity impact, but a transatlantic route that doubles capacity into petabit-class territory is a long-horizon infrastructure variable worth factoring into future assumptions about Europe-to-Americas bandwidth pricing and redundancy planning.
๋๋ฉ์ธ ์ง์ ์๋ ๋์์ด๋๊ฐ ํ์ ๊ธฐ์ค์ ๋ฐ๊พผ ๋ฐฉ๋ฒ
ํ ์ค
In Korea, every remittance service required users to pick a recipient's bank before sending money, treated as an unquestioned given -- until the Toss team, during user interviews, noticed people getting stuck on the account-entry screen because they couldn't find the bank they wanted to send to. This wasn't an isolated case: it showed up repeatedly across multiple users and was corroborated by a review of actual customer-support inquiries. A designer without banking or financial-domain knowledge questioned this seemingly obvious requirement, which led the team to reconsider whether users should have to manually select a bank name at all during a transfer, ultimately changing the team's standard. The piece frames this as a case where a non-domain-expert can add value in a mature, heavily regulated product precisely by challenging assumptions insiders have stopped questioning. This summary is limited to information confirmed through search-engine snippets, since the source page could not be opened directly.
๐ก A non-expert overturning a domain-expert assumption by validating it against real user interviews and support-ticket data is a useful argument for building cross-team review into engineering processes, since outsiders to a domain (e.g. onboarding or observability engineers) can surface friction insiders have stopped noticing.
Understand the top paths users take to convert or drop off with Journey Paths
Datadog
Journey Paths in Datadog Product Analytics visualizes the actual sequences of pages and actions users take through a defined workflow, ranked by frequency. It addresses a gap in traditional funnel analysis, which shows conversion percentages between steps but not what users actually did in between. The Converted Paths view surfaces inefficient routes โ users who repeatedly backtrack before eventually converting โ while the Drop-off Path view distinguishes two abandonment patterns: "dead ends," where a flow simply stops, and "detours," where users leave the intended path and never return. As a concrete example, Datadog cites a checkout funnel with 40% drop-off between checkout and payment, where Journey Paths reveals that many of those users click a support link and never come back โ turning an ambiguous metric into quantified behavioral data. From there, teams can pivot directly into Conversion Analysis (statistical ranking of attributes and segments correlated with conversion or drop-off), Session Replay (individual session recordings), or RUM (investigating errors, latency, and frustration signals), and the Journey Paths chart itself can be saved to a dashboard.
๐ก Because Journey Paths can pinpoint the actual behavior behind a drop-off (e.g., clicking a support link and not returning) rather than just a conversion-rate number, adding path-level behavioral views to observability dashboards can shorten the time from noticing a drop-off to fixing its root cause.
This digest was collected from RSS feeds and summarized by AI (Claude). See the original links for full details.