AMDGPU: understanding HDMI Passive VRR
AMD proposes keeping compatible HDMI displays in VRR signalling mode on the desktop. What the patches change, and why Linux 7.4 is not a delivery promise.
PeopleAreGeek / Publications
The journal
News, analysis and practical guides on development, networks and systems. The latest publications first.
Fable 5.1 Cache Pricing: Calculate the Actual Saving
Cache reads fall to $0.25 per million tokens. A worked cost example explains why that does not cut an entire Claude API bill by 75%.
Anthropic-Lambda: decoding the reported $35bn deal
A reported Anthropic-Lambda agreement points to Beacon Point. Distinguishing the reported tenant chain, contracted IT power and future delivery.
Empirik: evaluating an infrastructure change agent
Empirik announces funding and an agent that maps the impact of infrastructure changes. A practical way to test the claims before granting write access.
Linux 7.3: AMD ATL loads only when needed
The AMD ATL change avoids unnecessary module loading on Zen systems without enabled DRAM ECC. What this means for logs and memory-error monitoring.
Steam August 2026: understanding Linux’s 3.90%
Linux is 3.90% in Valve’s August survey. Recalculating the monthly change and CachyOS share without confusing percentage points with user counts.
Tripo P2.0: what quad topology actually changes
Tripo announces 3 billion yuan in funding and previews P2.0 quad output. What topology changes, and what still needs testing in a production asset.
WinBtrfs 1.10: snapshot fix and recovery limits
WinBtrfs 1.10 fixes snapshot writes and several Windows crashes. How to approach the update and why a Btrfs scrub is not a complete filesystem check.
Zen 6 CPPC: kernel constants are not CPU specifications
AMD’s CPPC proposal uses three fallback frequency values for selected Zen 6 clients. Their ratios explain scheduler preparation, not retail CPU performance.
Firefox 155: Happy Eyeballs v3 and QUIC v2 explained
Firefox 155 adds connection racing and QUIC v2 negotiation. What Mozilla confirms, what the IETF draft describes and what an endpoint operator can check.
KDE Linux: file history with Btrfs and KIO Snapshot
KDE Linux adds automatic home snapshots and access through Dolphin. How Btrfs, Snapper and KIO Snapshot fit together, and what they cannot protect.
Lekuo SouthBridge Max: more ports, one shared uplink
The B650-based PB65MX3 adds NVMe, SATA and USB through PCIe 4.0 x4. Its slot allocation and shared bandwidth are the limits to check before buying.
Nvidia Vera: why the scheduler prefers the same sibling
The Olympus SMT proposal prefers PE0 to avoid alternating active siblings. Both PEs have equal steady-state capacity; the reported gain is workload-specific.
Servo 0.5: faster canvas, with limits worth testing
Servo 0.5 reports faster canvas and text rendering. Why the headline gains are workload-specific, and how to evaluate an embedded application.
systemd 262-rc1: what LUO sessions preserve
LUOSession connects services to kernel live-update sessions. It does not transparently preserve every process, descriptor or device across a kernel replacement.
The Ur Project: a distro builder before a distro
Auke Kok describes a Rust distribution-building stack. Recipes, repositories and provenance have distinct roles; public readiness remains a separate milestone.
Incus 7.4: what near-live migration actually preserves
Incus copies container storage before stopping it for a final sync. How to plan the restart, estimate interruption and check ZFS or Btrfs prerequisites.
Linux 7.3-rc1: a large patch is not a speed score
AMD DCN6 headers explain much of Linux 7.3-rc1’s size. What the release message establishes and how to compare a test kernel with a known-good build.
Multikernel: several Linux kernels on one machine
The v7.0-mk2 tree partitions hardware between Linux kernels. Read its x86 support, vendor measurements and practical limits before comparing it with KVM.
OpenAI and Cursor: preparing for the proposed cutoff
OpenAI proposes November 12 for ending its Cursor model-supply contract. Map affected workflows and test an alternative without assuming all access ends.
OpenAI’s reported Mac fleet: environments versus GPUs
Reports of Mac purchases for computer-use training do not reveal where models execute. Separate desktop environments, inference and learning before sizing.
OpenShot 4.0: record separate sources, edit precisely
OpenShot 4.0 adds capture and grading workflows. Understand microphone versus screen audio, the macOS limitation and a short export check before a long take.
Pangolin 1.22: AI access rules and budget limits
Pangolin adds an AI Gateway and expands Community access features. Understand virtual keys, estimated budgets, unpriced models and transcript retention.
Alibaba Cloud Brazil: check services and failure domains
Alibaba announced its Brazil region on August 27. Two data centres do not automatically give an application zone-failure tolerance or every AI service.
DeepSeek funding report: what pre-money means
The reported 50 billion yuan raise and 500 billion valuation are different quantities. Calculate the implied post-money value without claiming a closed deal.
FFmpeg fruc_vulkan: interpolated frames and tradeoffs
The NVIDIA optical-flow filter creates intermediate frames. Check extension support, filter availability and the quality versus speed effects of perf and grid.
ROCm on FreeBSD: what vector addition would prove
The port is approaching a basic compute milestone. Understand compiler, runtime, driver and memory responsibilities without mistaking progress for support.
git.kernel.org: reading the scraper traffic estimate
The administrator estimates about 2% legitimate requests. Challenge completion is not proof of humanity, and request share is different from CPU cost.
Linux 7.3 KVM: three MMU responsibilities separated
KVM separates page-table format, guest walking and table building. Understand the cleanup, nested translation and the specific Xen timing fix.
Meta’s robot trials: a cable swap needs verification
Reported data-centre robot trials raise an operational question: how do you establish the correct target and verify recovery after a physical intervention?
x86 WBINVD hints: an allocation is not a CPU launch
Ludloff’s message documents cache-operation prefixes for an unnamed implementation. Read existing WBNOINVD behaviour, fallback and the limits of attribution.
a16z Machine Age: $1.1B for AI infrastructure
The Machine Age Fund targets AI hardware. What the announcement establishes, and why more compute does not automatically remove a system bottleneck.
Anthropic MHS: how agents reach physical devices
MHS is a research preview for programmable instruments. QuEra’s example separates AI-assisted controller development from deterministic execution.
Btrfs: Chris Mason moves to a reviewer entry
Chris Mason’s Btrfs entry changes from maintainer to reviewer. The patch updates contacts and records a handover he says happened years earlier.
Debian’s AI vote: responsibility stays with authors
Debian selected Responsible Use of Generative AI. Disclosure is encouraged, while review, licensing and responsibility remain part of contributing.
Microduck: Open Software, Hardware Still to Evaluate
Microduck opens pre-orders at $399. Its press kit distinguishes open software from hardware and lists the specifications that remain provisional.
Linux 7.3: what the USB-C cable checks really test
The new USB-C helper distinguishes supported, unsupported and unknown cables. Missing identity data does not automatically mean the cable is rejected.
ReactOS 0.4.16: setup, Server Core and Wine limits
ReactOS combines its live and installation image. Wine 10 synchronization is still underway, and Server Core keeps the Win32 subsystem.
Flatpak’s €508,640 funding: permissions on the roadmap
The funded Flatpak work targets finer permissions and portal maintenance. Planned features should not be confused with protections already installed.
ROCm 10: agent tools and the limits of speedup claims
ROCm 10 adds ROCm.AI, including preview tools and Hyperloom. Evaluate optimization with correctness checks and end-to-end timing, not kernel speed alone.
Asahi Linux on M3: released, with important limits
M3 support reached Asahi’s installer on September 6 in Expert mode. GPU acceleration, sleep, MacBook HDMI and the M3 Ultra still need careful distinction.
Huawei and HP: what the Wi-Fi patent deal establishes
The companies announced a global patent cross-license. The public statement does not establish the performance or Wi-Fi certification of any product.
Linux 7.3 MGLRU: protect code, read timings correctly
MGLRU promotes mapped executable folios after first use. The reported 15% reduction concerns system CPU time in a constrained build, not total elapsed time.
Linux 7.3: RWF_DONTCACHE reaches block devices
Block-device I/O gains dropbehind support. It still uses the page cache, and pruning is conditional rather than an immediate guarantee after every call.
WebMCP site tools: a concrete read-only example
Register an article-outline tool with WebMCP, understand ChatGPT’s supported subset, and keep annotations separate from actual permissions.
Ubuntu 26.04.1: release media and the upgrade gate
Ubuntu 26.04.1 media are available, but normal upgrades from 24.04 were deliberately delayed. A point release does not restart the support clock.
Anthropic and Nscale: what the $45bn report establishes
The reported Anthropic contract concerns future capacity. Separate the tenant allocation, the Monarch campus plan and the power needed to operate it.
DuckLabs joins AWS: what changes for DuckDB users
DuckLabs says its AWS transaction concluded on August 31. DuckDB keeps its MIT license and Foundation stewardship; deployment choices remain separate.
AWS plans 2 million more Nvidia GPUs for 2027-2028
AWS and Nvidia plan additional GPUs across several generations. The announcement is a capacity roadmap, not an instance availability or performance result.
Linux 7.3 ksmbd adds opt-in Time Machine support
The kernel SMB server gains Apple interoperability. A visible share is only the start: verify backup, recovery and reconnect behavior on the intended setup.
Linux 7.3 NFSD: notifying a cache instead of recalling it
CB_NOTIFY lets cooperating NFS clients receive directory changes. The benefit depends on delegation and notification support, not just a server upgrade.
Kernel Zstd: why the BMI2 patch helps small VM requests
The revised Zstd proposal uses kernel CPU feature dispatch. Its large decompression gain comes from a specific 4 KiB KVM benchmark, not all workloads.
Nvidia confirms its agreement to acquire Hugging Face
Nvidia confirmed a $12.93bn agreement on September 3. Its platform commitments should be distinguished from each model’s license and actual deployment needs.
wmbench: compare desktop speed without confusing watts
wmbench separates uncapped rendering from fixed workloads. Backend, window geometry, missing protocols and sensor coverage determine what a comparison means.
Mechanical Turk closes September 30: dates to separate
MTurk’s official closure FAQ separates task submission, approvals, bonuses and transaction history. Map each workflow to the relevant deadline.
Cisco’s AI Factory expansion: two fabrics, one offering
Cisco adds Supermicro systems to its AI portfolio with orders planned for October. Front-end and back-end networks retain different traffic roles.
Emerald AI raises $150m: flexibility is not energy saved
Emerald AI funds power-flexible data centers. Reducing a grid peak and reducing total energy are separate outcomes that need different measurements.
Granite 4.2: thinking modes and the limits of a 3B model
IBM’s 3B, 8B and 30B models share thinking controls, but not every training stage. Model weights and usable context require separate memory budgets.
MetaRoCE: tolerate packet loss without losing the data
MetaRoCE moves recovery and path control to endpoints. Its August description reports tests and schedules public specification and tooling for October.
Groq 3 LPX production: token speed is part of an agent
Nvidia says Groq 3 LPX is in full production. Its reported Gemma 4 result concerns token generation; tool waits and prompt processing remain separate.
Jalapeño benchmarks: read the power denominator first
OpenAI’s Jalapeño results normalize throughput by rated accelerator power. That is different from measuring complete-system energy for a matched workload.
![Original matrix-vector example: stored rows [1, 2] and [3, 4] multiplied by input [5, 6] give 17 and 39. Bringing input to arithmetic near stored rows can reduce weight movement. Not Samsung’s exact layout or instruction sequence.](/media/samsung-lpddr5x-pim-hot-chips-2026-explained-en.webp)
Samsung LPDDR5X-PIM: internal bandwidth is not host I/O
Samsung’s PIM design computes beside DRAM banks. Its internal bandwidth and demonstrated token rate measure different parts of the inference system.
Mac Studio M5 Ultra: What 512GB Changes for Local AI
Apple announces a Mac Studio with up to 512GB of memory. A sizing calculation explains what capacity tells you, and what still needs a benchmark.
ath11k: lower loaded latency comes with a throughput cost
The ath11k v3 proposal enables airtime queue control. Its 22 ms result uses tighter settings and lower throughput; defaults do not deliver that gain.
Infineon-C2i: why AI power delivery needs better control
Infineon’s planned C2i acquisition concerns digital power control and future packaging. A simple current calculation explains the distribution problem.
Intel Xe’s black-screen fix: round the VRAM limit down
The Xe fix protects compression metadata by rounding usable VRAM down to 4 KiB. Its changed assertion matters too; this was more than a word swap.
Wan3.0: API access, reference inputs and available weights
Wan3.0 supports up to 30-second video and multimodal references. Check input combinations, region and model weights separately from Alibaba’s financing.
Hugging Face: make an offline model startup reproducible
The sale report now has an official follow-up. Use pinned revisions, complete artifacts and an offline startup rehearsal to assess dependence on the Hub.
IBM Plans Arm and Z Instructions in the Same Core
IBM presents a future dual-architecture processor. Native instruction support is significant, but it does not settle application or operating-system support.
Linux 7.3 bounds EFI waits, then disables runtime calls
A 120-second EFI completion timeout contains a firmware hang. It disables runtime services rather than repairing firmware or making later EFI writes work.
Linux 7.3 prepares DAX; FAMFS remains a separate proposal
Merged DAX fixes prepare for FAMFS. Its separate proposal maps shared memory as files, without making all CXL memory persistent or automatically coordinated.
Omacom reaches $15.5m: cash, pledges and tokens differ
Omacom’s updated total includes token credits and multiyear commitments. Its Hyprland sponsorship starts October 10; the initial $8m headline is outdated.
ReactOS Job Objects: process groups, with limits to scope
PR 7500 adds basic pre-Windows-8 job support. Closing any handle does not automatically kill every process, and the merge does not prove browser isolation.
AMD RDNA 4m firmware: what reached Linux, and what did not
The August 21 firmware merge adds GC 11.7 files and associated display and security firmware. It does not certify a complete Linux graphics experience.
Anthropic hires Amir Salek: chip design, not chip delivery
Anthropic confirmed the hire of former Google TPU leader Amir Salek. Its co-design effort is real, but the report supplies no finished accelerator or ship date.
Plasma 6.8: what the 50 ms panel delay actually changes
KDE reduces a panel visibility timer from 500 to 50 ms. The change concerns waiting after pointer movement, not a tenfold speedup of the desktop.
Nvidia server price report: read the memory assumptions
Reported increases above 15% concern selected AI server configurations shipping in early 2027. A weighted-cost example explains why memory content matters.
GPT-5.6 Sol at $4/$20: calculate the actual saving
Sol standard API pricing is $4 input and $20 output per million tokens. Calculate a mixed workload and account for long-context and promotional conditions.
AMD UALink patches: pinned memory and authorized imports
The August 31 revision clarifies AMD’s UALoE proposal: exported memory is pinned and remote access uses authorized handles. Upstream delivery remains separate.
Apprise 1.13.1: keep updates controlled before 2.x
Apprise 1.13.1 followed the planned final 1.x release with attachment fixes. Bound the major version, lock tested dependencies and validate real delivery.
GitHub’s August outage: the sidecar and the retry loop
GitHub’s incident report links sidecar saturation to cascading authentication failures. A retry calculation explains why restoring capacity was not enough.
Gemma’s billion downloads: what the number can tell you
Google reports a billion Gemma downloads and over 100,000 variants. Download events, unique devices and successful deployments are different quantities.
NetworkManager 1.58.1: isolate the WPA3 auto-connect fix
The SAE auto-connect fix concerns wpa-psk profiles. Separate that path from routing and DNS failures, and account for stricter private 802.1X profiles.
Nvidia and Poolside: license fee, equity and Model Factory
The reported $6 billion license and $1 billion investment are different transactions. Poolside’s technical writing explains the training system behind the story.
OpenZFS 2.4.4: what MMP reclaim actually recovers
The new recovery command addresses unreachable mirror legs still expected by the pool. It preserves activity checks and requires a fenced failed peer.
Wine 11.16: what its VA-API H.264 path requires
The VA-API backend adds a Direct3D video decode path for H.264. Application use, host driver support and the Wine build all matter; gains require measurement.
Azure VMware licensing: which deadline applies to you?
PayGo and reserved license-included AVS capacity have different deadlines. Check reservation expiry, portable VCF registration and the actual BYOL core count.
Antigravity: editor extensions and enterprise support
Check the editor, sign-in route and enterprise preview status separately. Google’s official settings demo shows where administration controls are configured.
Netdev’s AI patch pressure: review capacity is the issue
Jakub Kicinski describes growing review pressure in the Linux 7.3 pull request. His AI estimates and future automation plans need to retain their qualifiers.
Rust 1.98 algebraic floats: accuracy and regrouping
Algebraic methods permit regrouping without promising a universal speedup. A binary64 example shows the numeric effect; Rust 1.98.1 also fixes a separate bug.
Btrfs in Linux 7.3: where the reported gains apply
Direct I/O, metadata tracking and fsync improve in specific cases. Read each measurement separately and check the changed cache and recovery mount behavior.
ChatGPT ads in Europe: plans, controls and access
European advertising has expanded, with Ads Manager self-service announced August 31. Personalization controls and the limited Free Ads-Free option differ.
Go 1.27 JSON: keep the API or choose stricter v2?
The old JSON import uses a new engine while preserving semantics. Explicit v2 has stricter defaults; a migration matrix and Go 1.27.1 clarify the next step.
Linux 7.3 scheduling: flattening the task choice
The EEVDF task choice is flattened within each CPU’s fair runqueue. Cgroup accounting remains hierarchical, and a September correction fixes throttling.
Linux 7.3 memory locks: what the benchmarks measure
KSM’s 705-to-1.67 ms result is a specific lock-hold benchmark. zsmalloc improves another path, with architecture and configuration limits on lock removal.
Marvell and Google: decode the $12.2bn warrant
The filing describes conditional share rights, not a $12.2bn cash gift. Vesting, chip purchases and cash or cashless exercise are separate steps.
Veeda AI’s $90m seed: what a world model must prove
Veeda’s funding supports learned simulation for physical AI. A branched-action example explains why realistic video alone cannot validate a robot trainer.
Cerebras CS-4: read the wafer and rack numbers
Three WSE-3T wafers provide the advertised 750 sparse-FP16 PFLOPS. Memory bandwidth, external I/O and measured token speed describe different things.
Cursor Origin: what a GitHub mirror really changes
Origin can host code or mirror GitHub, but mirroring keeps GitHub authoritative. Check pull requests, CI, permissions and sync before treating it as failover.
Etched: what its first Jane Street rack establishes
Etched raised $700 million at a $21 billion valuation and shipped a first customer rack. Delivery, customer contracts and independent benchmarks are distinct.
Europe’s data centres: what the 175 km average means
JLL compares planned hyperscale sites with an earlier delivered cohort. Distance affects network design, but the average is neither a relocation nor an RTT test.
Nvidia H200 in China: shipments versus deployment
Nvidia’s quarterly filing confirms limited H200 shipments. Reported ByteDance and Tencent quantities remain separate from company-confirmed deployment data.
Samsung Foundry: why a 15% rise is not one price
Reported wafer-price increases vary by node and customer market. Good-die yield, packaging and contract scope determine the cost of a finished chip.
Warp Factories: evaluate the workflow before scaling
Warp’s cloud-agent platform now includes Factory Benchmarks. Compare configurations on fixed tasks, count retries and keep correctness checks independent.
AI commitments: what the $3 trillion headline means
Future contracts, cash capex and working capacity are different measures. Alphabet’s filing shows how to read obligations without turning dollars into GPUs.
Fedora 45: why the extra x86-64-v3 builds were rejected
FESCo rejected additional v3 packages for Fedora 45 on August 17. Resubmission needs performance and image-planning evidence; Fedora 46 is not guaranteed.
Firefox 154: iterator edges and local-network access
Understand chunks versus windows, one-based CSS sibling indexes and local WebSocket permissions. Firefox 155.0.1 now fixes subsequent sidebar and CSS issues.
Groq’s $350M round: funding is not deployed capacity
Groq targets more than 200 MW in 2027 and plans Nvidia LPX deployments. Financing conditions, installed power and useful inference throughput remain distinct.
Linux 7.3 FailFS: what it blocks and what remains
FailFS removes ambient path access, but retained directory descriptors still matter. Check entry conditions and distinguish retired SGI EFS from Amazon EFS.
Wispr Canto: word errors and clean dictations differ
Wispr announced a Canto preview alongside its $280 million round. Its accuracy claims need context, and zero-edit rate cannot be inferred from average word errors.
Higgsfield raises $400M: reading its growth figures
Higgsfield announces a $5.4B valuation and $700M annualized revenue. What those figures measure, and how to evaluate the video platform beyond generation counts.
Imagen 4 retired: migrate the whole image contract
Google lists August 17 for Imagen 4 retirement. Migrate model selection, requests and image parsing; Firebase and the direct Gemini API use different examples.
Linux 7.2: cache locality and USB4 byte streams
Linux 7.2 adds cache-aware scheduling and USB4STREAM. Understand the transfer path, configuration requirements and late DRM scheduling reverts.
Microsoft AI capacity: why gigawatts need context
A disputed chip estimate raises a useful question: what does installed AI capacity measure? Read Microsoft’s capex figures and power boundaries correctly.
NVIDIA’s $105B Ohio guarantee: what it covers
PORTS-Pike combines an OpenAI lease, SB Energy construction and NVIDIA credit support. The guarantee ceiling is neither immediate spending nor completed capacity.
Stripe and OpenRouter: agreement, routing and billing
Stripe confirmed an acquisition agreement on August 19. OpenRouter says integrations stay unchanged; understand what routing and billing each control.
Anthropic’s Q2 revenue: interpreting the reported jump
Preliminary revenue reporting needs careful arithmetic. Separate quarterly sales, annualized pace and adjusted operating income before drawing conclusions.
Debian’s LLM ballot: eight proposals, nine choices
The August vote is complete. Understand the default option, majority thresholds and why the Social Contract ban did not fail solely because of a 3:1 rule.
NVIDIA’s 13F: holdings value is not fresh spending
The June 30 filing shows large Intel and SpaceX positions. Compare share counts, reported value and purchase cost without treating the table as a trade history.
SpaceX owns Cursor: what changed after closing
The August 14 close is confirmed in a filing. The later OpenAI supply notice matters more to users than unverified forecasts about branding or default models.
Apple’s proposed link-out fees: what 15% means
Apple proposed US link-out commissions in court. A worked example separates the requested rate, processing costs and checkout conversion.
Gemini watermarks: three signals, one visible switch
The visible watermark toggle does not remove SynthID. Understand the separate roles of on-screen branding, embedded signals and Content Credentials.
Linux 7.3: why equal E-core clusters need balancing
Intel’s scheduler corrections address load stranded across equal-capacity clusters. See the mechanism, current source and limits of performance claims.
Copilot app merger: which content actually survives?
Copilot’s update treats ordinary chats, group chats and saved research differently. Check the migration path before assuming every feature carries over.
Qwen3.8-27B: open weights do not settle memory needs
Qwen3.8-27B offers downloadable weights, but precision, context and serving overhead determine whether a configuration actually fits your hardware.
RustDesk on Wayland: how unattended capture works
RustDesk’s Wayland preview uses a privileged DRM capture path. Understand the service boundary and what to verify before relying on remote access.
Blacksmith’s $45M round: measure the CI bottleneck
Faster runners help only part of a CI pipeline. A worked example separates queueing, execution and fixed steps when assessing Blacksmith’s funding story.
Databricks’ $5B round: money, run-rate and agent state
Databricks links its funding to agent infrastructure. Separate valuation from revenue, and database state from proof that an external action completed.
DeepSeek V4 Pro pricing: calculate the whole request
DeepSeek’s current peak schedule applies on weekdays. A token-mix calculation shows why the cache-price headline is not the increase in every bill.
Gemini 3.7 Flash: budget beyond the introductory rate
The 3.7 Flash launch remains historical news, while 3.8 is now available. Calculate current and scheduled token charges without hiding thinking costs.
Kioxia’s 2Tb QLC: chip density is not SSD capacity
The ninth-generation NAND announcement combines CBA and six planes. Decode its bit units before confusing chip density with usable drive capacity.
OpenAI Ultrafast: 14× tokens is not 14× every task
Ultrafast previews faster GPT-5.6 Sol inference on Cerebras. A latency breakdown explains why token throughput and completed-task time differ.
Podman 6.1: volume rename has stricter preconditions
Stopping a container is not enough to rename its volume. Podman 6.1 also clarifies image-declared storage through Quadlet’s ImageVolume setting.
RDCMan’s RDG changes: preserve a real rollback copy
The August Sysinternals update included RDCMan 3.20; the current release is 3.21. Shared RDG files need both format and credential compatibility checks.
Claude watermarks: what detection can actually tell you
Claude’s detector is in private preview. Understand supported models, weak signals after proofreading, and why a watermark does not establish authorship.
CodeRabbit’s $143M round: measure review usefulness
CodeRabbit expands into change triage and security. A worked example shows why review volume and alert count cannot establish defect-detection quality.
Comma Chestnut: two USB paths, different requirements
Chestnut costs $249 as a dock; the complete kit now lists at $899. Separate USB4 docking from custom USB3 compute and check the TinyGPU requirements.
Gemini’s billion users: read the right denominator
Google reports one billion monthly Gemini users. Voice users, Live interactions and image output measure different things; a simple example explains why.
IBM and Together AI: what the $240M agreement covers
The multi-year agreement targets NVIDIA infrastructure on IBM Cloud, expected in early 2027. It does not establish an equity investment or a GPU unit price.
Lovable’s $400M round and the limits of drafts
Lovable raised at a $13.3B valuation. Its new drafts isolate frontend exploration, but currently use the published app’s database: understand that boundary.
Manjaro 26.1: an ISO snapshot, not a reinstall rule
Bian-May now has 26.1.1 images. Understand desktop and kernel variants, rolling updates, and why an installed kernel is not necessarily the running one.
MacBookPro15,1: why waking the second GPU matters
The Apple GMUX patches target a specific Intel MacBook Pro. Understand the GPU power cycle, audio dependencies and the limits of the reported idle saving.
LACT 0.10: read the sensors before changing controls
LACT adds NVIDIA controls and fixes Blackwell hotspot readings. Separate telemetry from tuning, and hook configuration authorization from hook execution.
Linux 7.2 kept FIFO for the shared DRM scheduler
The fair-policy default was reverted before Linux 7.2 shipped. Check the final source, and separate GPU scheduling from dependency waits and CPU scheduling.
Linux GUP batching: where the reported 12.8× comes from
The RFC reports a large gain for 64 kB mTHP, while 2 MB THP barely changes. Read the shared baseline and distinguish page lookup from application speed.
Mold 2.42: correct metadata matters as much as speed
SFrame v3, Android relocation packing and ICF fixes change more than link time. A small address example shows why surviving code still needs valid metadata.
NVIDIA’s $500B financing plan is not cash already raised
The announced platforms aim to mobilize third-party capital under final agreements. Separate that target from funded projects and the value of collateral.
ChatGPT on Linux: packages, preview and local projects
Choose the right Linux package and architecture, understand XWayland defaults, and separate local project access from running an AI model offline.
QEMU 11.1: test byte order, not imaginary hardware speed
UFS emulation and big-endian RISC-V improve functional tests. Understand the byte-order example and the host requirements behind nested HVF virtualization.
Anthropic and Theseus: who finances the buildings?
Macquarie and GIC plan infrastructure leased to Anthropic. A financing example explains why majority equity does not mean most of a project’s total cost.
FastFlowLM 1.0: detecting an NPU is only the first step
FastFlowLM joins ROCm and adds SmolVLA. On Linux, understand the separate kernel, XRT and model requirements before interpreting a successful validation.
Fedora on Panther Lake: check the profile provider
One Framework laptop improved with power-profiles-daemon. Understand why identical profile names can hide different settings and how to compare fairly.
CoreOS, zram and oomd: three memory policies to check
The Fedora 45 CoreOS plan combines compressed swap and oomd. Kubernetes startup, Pod swap policy and pressure-based termination still need separate checks.
Nova: what TLV firmware and vGPU boot actually add
Nova’s firmware format carries tagged metadata. Understand its alignment and validation rules, and why GSP boot is only one part of virtual GPU support.
OpenAI’s reported buyback: follow the cash direction
Bloomberg reports a $7 billion employee share repurchase at an $852 billion valuation. Why a buyback differs from fundraising and does not prove retirement.
Raspberry Pi AXI PMU: measuring shared bus activity
Proposed perf support covers Raspberry Pi interconnect monitors, including Pi 5. Understand what shared counters reveal and what they cannot attribute.
YouTube YPP 2027: entry and Shorts earnings differ
YouTube’s February 2027 changes distinguish new entry from continued Shorts earnings. Check the January terms deadline and understand rolling view totals.
CachyOS August: server profiles are an early step
CachyOS adds experimental server installation profiles and rewrites tools. Understand what changes for existing installations and what server readiness needs.
Intel’s share offering grew beyond the $15B headline
Intel priced a $20 billion offering at $95, then underwriters exercised the extra option. Follow the share counts, gross proceeds and dilution arithmetic.
Linux 7.2-rc7: more fixes do not measure AI accuracy
Torvalds linked rc7’s size to AI-assisted review. What the announcement says, what commit totals cannot prove, and how to make a useful regression comparison.
Linux Mint previews kernel retention by series
Mint’s planned kernel manager distinguishes tracked, custom and protected kernels. Understand the cleanup policy and why environment changes need new sessions.
Mesa Large GRF: the change is in Intel’s executor
The August Mesa patch enables 256-register tests in Intel executor and adjusts validation. It does not itself establish a new automatic policy for all games.
Muse Glimmer: budget for more than the model weights
Meta’s local agent model has separate text, vision and draft files. Check runtime support, context allocation and total memory before choosing a GGUF build.
Agent Plugins 1.0: what actually travels between agents
A common manifest can move skills and MCP configuration between clients. Transport support and execution permissions still need separate verification.
Cloudflare: machine traffic does not equal origin load
Cloudflare reports a non-human majority on its own network. Understand the measurement boundary and why more edge requests need not mean more origin work.
WeatherNext Cyclones: choose the right checkpoint
DeepMind opens cyclone forecasting code and weights. Training cutoffs, Mini resolution and ensemble interpretation determine what a reproduction can show.
Dell 7320 Detachable: a frame is not full webcam support
Linux patches investigate the 7320 Detachable camera. Sensor identity, reset wiring and RGB-IR pixels explain why capturing a frame is only one milestone.
Genesis Open Models: applications are not checkpoints
DOE’s GS1 initiative solicits scientific contributions. The official portal separates application from data delivery; a model release needs its own evidence.
Linux 7.3: why equal-capacity clusters still balance
Scheduler changes distinguish movement between equal-capacity clusters from moving a task to a stronger CPU. The distinction does not promise a universal speedup.
OpenAI and NextSlide: what editable slides really mean
NextSlide confirms its team joined OpenAI to work on ChatGPT. Distinguish the acquisition statement from product promises and genuinely editable slide content.
AMD and Taalas: decode speed is only part of latency
AMD plans to acquire Taalas. Its specialized HC1 silicon demonstrates fast Llama generation, but model quality and end-to-end response time still matter.
Linux on M3 Pro, Max and Ultra: boot is a first stage
Linux 7.3 gains initial M3 Pro, Max and Ultra device descriptions. Check Asahi’s per-device matrix before confusing a boot framebuffer with desktop readiness.
ByteDance’s reported 10T model: what the number means
The reported ByteDance training target is not a released specification. A parameter count can support a storage calculation, but cannot rank model quality.
Lumilens: the shipping claim concerns scale-out optics
Lumilens reports over $900 million raised and a first scale-out product shipping. Its broader NPO/CPO roadmap needs separate deployment evidence.
NVIDIA NOOA: a typed answer can still be wrong
NOOA puts agent state, methods and prompts in Python classes. Typed results help integration, but business rules and execution isolation remain separate.
Triton graphics now has an experimental UTM beta path
Triton connects Windows graphics to Neptune and host renderers. UTM 5.0.5 beta supplies an experimental path, with guest tools and backend limits to check.
Linux TTM: GPU memory protection has two strengths
TTM tries eviction before fallback for protected allocations. Understand min versus low, cgroup hierarchy and why protection is not unlimited physical capacity.
Wine 11.15: IStream support and the washed-out colors fix
Wine 11.15 lists an MSXML3 IStream fix and double-sRGB correction. An original calculation explains the color error without pretending to be a game screenshot.
AMD Spur: Slurm compatibility needs a workload check
Spur combines Rust, GPU placement and Slurm-style commands. Evaluate job semantics, accounting and controller failure before treating it as a replacement.
Btrfs fixup: dirty memory needs writeback preparation
Linux restored Btrfs COW fixup before 7.2 final. The merged code explains the missing reservation path and why a clean scrub cannot prove every write arrived.
GCC 16.2: a bug-fix release can change generated code
GCC 16.2 fixes more than 102 bugs. Compare it with 16.1 using the same language mode, flags and dependencies before removing any workaround.
OpenAI speaker report: the missing privacy specification
Bloomberg describes a screenless OpenAI device. Its reported shape and price do not answer what it records, sends to the cloud or retains.
Page Alloc Hogger: free bytes do not prove a block fits
The Linux RFC targets memory by node, zone, order and migration type. A simple page diagram explains why free capacity and contiguous capacity differ.
PoCL 7.2 final: what RISC-V CPU conformance establishes
PoCL 7.2 is now final. Khronos lists specific Star64 and Milk-V CPU configurations; that does not certify the board GPU or predict application speed.
CXMT in laptops: identify the part, not just the brand
Nikkei reports limited CXMT adoption by HP, Asus and Acer. Firmware inventory can help trace a machine, but it does not certify its DRAM dies or reliability.
Discovery Loop: faster experiments still need a holdout
Jeff Dean and colleagues founded Discovery Loop to automate research. Its ambition makes experiment selection and independent evaluation central questions.
Mesa 26.2 and NVK: mesh shaders need feature checks
NVK gained mesh shaders in Mesa 26.2. Query the actual GPU features and enable them; the Mesa version alone does not establish what an application can use.
Muse Code: worktrees do not isolate every side effect
Meta’s terminal agent supports persistent work. Check workflow availability, shared services and the chosen data tier before evaluating its output.
Qualcomm Synx: the dma-fence contract is the obstacle
The Synx RFC proposes cross-processor synchronization. Later review challenges its dma-fence interoperability because completion must respect memory-management rules.
Rust Coreutils: a test percentage is not full parity
Version 0.10 passed 645 of 690 listed GNU tests. The newer 0.11 uses a different tally, while rm --one-file-system was still unimplemented in 0.10.
USB4STREAM busy polling changes the application contract
Linux 7.3-rc2 contains busy polling in stream reads and writes. It changes poll behavior and can consume CPU while waiting; it is not a universal latency switch.
AMD’s eSPI RFC: four channels, one shared link
AMD proposes a Linux eSPI bus framework. Its first RFC implements link configuration and reset, while channel operations and interrupt handling remain deferred.
Apple versus OpenAI: what the injunction request says
Apple’s motion concerns alleged confidential information and evidence. OpenAI disputes it and links its September opposition; neither filing is a judgment.
C-Kermit 11: test the transfer policy before migrating
C-Kermit 11 changes collision and transfer-mode defaults. A concrete migration matrix checks the destination, payload and failure handling; 11.0.509 is available.
Cloudflare OS: sharing an app is not copying its data
Cloudflare OS separates collaborative gadgets from reusable blueprints. Its early-access architecture needs permission and sharing tests before company deployment.
FEX 2608: a narrower WFE fix, plus the 2609 update
FEX 2608 changes Wine thread-suspension waits, not every x86 spinloop. Version 2609 adds an optional JIT disk cache with growth and invalidation caveats.
Intel DEC on Linux: zero means no desired target
An intel_pstate change clears HWP’s desired-performance field when DEC is enabled. It keeps the other controls and is present in the Linux 7.3-rc2 history.
Tokenomics: calculate cost per accepted result
The new Tokenomics Foundation plans common AI cost methods. A worked example explains why cheaper requests can produce more expensive usable outcomes.
Linux process builder: the API precedes the backend
Li Chen’s RFC creates a taskless pidfd before launching. Its implementation still uses vfork-style machinery; a new interface is not yet pristine process creation.
Volta’s Norway deal: 133 MW gross, 121 MW for IT
The reported Anthropic compute deal sits above a separate Bitdeer lease. Power boundaries and contract scope matter more than a single headline megawatt figure.
FFmpeg 9: animated WebP needs timing checks too
FFmpeg 9 adds animated WebP input. Its libwebp_anim encoder and WebP muxer also exist; a useful conversion must preserve timing, loop policy and transparency.
GNOME Boxes Beta: How to Assess the Flatpak Rewrite
The GNOME Boxes rewrite bundles its virtualisation stack. Here is a concrete evaluation plan that keeps beta testing separate from existing virtual machines.
Linux staging’s LLM policy: scope and exception
Greg Kroah-Hartman’s August 3 policy concerns drivers/staging. Its security-fix exception requires actual hardware tests and a defensible report, not just a build.
NVIDIA funds LVFS: check firmware support per device
NVIDIA’s LVFS sponsorship accompanies DGX Spark firmware rollout. A sponsor name does not establish update availability for every device carrying the brand.
Qwen3.8 weights are out: Max and 27B differ
Qwen3.8’s released 2.4T weights are text-only and require thinking. The 27B model supports vision; licensing and memory needs differ from the hosted Max service.
Wild 0.10: measure linking and debugger startup together
Wild adds --gdb-index and changes output handling on Btrfs/vfat. An index moves work to link time; debugger reuse and cache state determine the useful tradeoff.
AerynOS 2026.08: updates that can update themselves
Versioned repositories let moss cross on-disk format changes. The August ISO incorporates that transition; reflinks and network naming need separate attention.
GNOME 51 blur: the app asks, the compositor decides
Mutter 51 beta adds Wayland background blur; the release candidate adds policy. A requested region still needs client support and a readable fallback.
Kakehashi: useful experiment, maintenance has stopped
The macOS-to-Linux ARM64 layer now documents git and clang, but its maintainer has stepped away. Compatibility, host access and support remain separate questions.
Linux 7.3 and Realtek 10GbE: check the whole port
RTL8261C/D support is present in Linux 7.3-rc2. The PHY driver, firmware, MAC and board integration all matter before a port can deliver useful 10GbE throughput.
MacBook Air shortages: a dated report, not live stock
August reporting described constrained MacBook Air supply. Server-memory forecasts provide context, but cannot establish today’s delivery date for your configuration.
Windows 11 memory work: measure the job, not free RAM
Microsoft’s July quality update targets memory efficiency on PCs with 8GB and above. The useful test is responsiveness under the same workload, not an idle screenshot.
Box64 0.4.4: what the default code cache can save
DynaCache reuses translated code, with a default disk limit and a DIRTY-mode exclusion. The configurator helps inspect profiles; faster starts still need measurement.
AI Act Article 50: marking and disclosure differ
The transparency rules apply from August 2, with a limited marking transition in Commission guidance. Identify the actor, output and exception before adding a label.
GNOME RFCs and extension quality are separate tracks
A draft RFC framework would preserve project decisions. Extension guidance already asks authors to understand and maintain code, including its cleanup lifecycle.
Go collections: decide what makes two keys equal
The Go1.28 collection umbrella remains a proposal. Go 1.27 supplies Hasher; matching the hash to the equality relation is the essential design constraint.
Intel Starfire: EDAC support is not space qualification
Linux recognizes Starfire’s memory-controller ID. Intel’s newer brochure lists up to 80 total TOPS, while radiation characterization remains explicitly in progress.
Linux’s Zen 5 fix: eight model IDs, not eight products
The correction assigns Family 1Ah models D0-D7 to Zen 5. Family/model classification is distinct from marketing names, shipping products and measured performance.
NetBSD 11: a 10ms kernel is not a 10ms service
NetBSD 11 adds stable RISC-V support and a lean x86 MICROVM kernel. The project’s boot figure excludes the separate question of when your application is ready.
Dasharo 0.9.0: check the board, CPU and recovery path
The MSI PRO B850-P WIFI release supports Phoenix, not every AM5 processor. Power restoration and update limitations need a more precise reading.
GEMA v Suno: what the Munich ruling actually establishes
The July 31 ruling concerns six musical works and reproducible outputs. Its reasoning does not turn every AI model into a proven copy of every training work.
Google Earth rollback: a mapped scene is not evidence
Google withdrew its Earth image-generation feature after launch. Generated scenes were not public map replacements, but screenshots could lose their context.
KDE multi-GPU work: when an eGPU avoids extra copies
The fast path depends on direct scanout to a display attached to the rendering eGPU. A KWin change alone does not make every application or driver eligible.
Linux SATA quirks: match the WD model before acting
Linux disables link power management for specific Western Digital model strings. A targeted quirk is different from disabling it for every SATA device.
Servo 0.5: page rendering is only part of compatibility
Servo 0.5 follows July’s 0.4 report with broader platform binaries and more fixes. A page that looks correct still needs input and state checks.
EU Opens Bidding for Seven AI Gigafactories
The European Commission opened its AI Gigafactories call on July 30, 2026: up to seven sites, 75,000 to 100,000 chips each, bids close November 12.
Gemini Robotics 2: Three Models, One Runs on the Robot
Google DeepMind shipped Gemini Robotics 2 on July 30, 2026: a whole body action model, an ER 2 reasoning model in preview, and an on-device variant.
Nouveau Gains NVDEC Channels, Unlocking Vulkan Video in NVK
A Nouveau patch queued for Linux 7.3 lets userspace allocate NVDEC channels through ABI16, which is what Mesa's NVK driver needs to advertise H.264 decode.
Nscale Buys Anyscale, the Company Behind Ray
Nscale agreed on July 30, 2026 to acquire Anyscale, adding an orchestration layer above its GPU fleet. Ray itself stays governed by the PyTorch Foundation.
OpenAI Cuts GPT-5.6 Luna 80% Three Weeks After Launch
OpenAI cut GPT-5.6 Luna API prices by 80% and Terra by 20% on July 30, 2026, three weeks after the family reached general availability. Sol is unchanged.
OPI Ships Its First Release to Standardise DPU and IPU APIs
The Linux Foundation's Open Programmable Infrastructure project published OPI Abstraction v0.1.0, a vendor neutral API layer across 26 repositories.
Amazon Leo Files for 5,105 Direct to Phone Satellites
Amazon asked the FCC to authorise up to 5,105 satellites for direct to device service on Globalstar L-band and S-band spectrum, with deployment starting in 2028.
AMD Patches Bring Ryzen Master Style Controls to Linux
AMD posted kernel patches on July 29, 2026 extending the amd_hsmp driver to Zen 6 client chips, using the Ryzen Master SMC message set instead of the EPYC one.
ChipAgents Raises $60M to Point Agents at Chip Verification
ChipAgents announced $60 million on July 29, 2026, led by B Capital, taking its Series A to $131 million across three tranches, with agents aimed at verification.
UK Regulator Probes Microsoft 365 Copilot Price Rises
The CMA opened an investigation on July 29, 2026 into whether Microsoft gave 365 Personal and Family subscribers clear information about the Copilot price rise.
GlobalFoundries Gets $300M to Move Optics Into the Package
A $300 million CHIPS award funds GlobalFoundries work on co-packaged optics, targeting 400Gb/s and a 5x energy efficiency gain, with Washington taking 1% equity.
Azure Passes $100B a Year and Still Runs Short of Capacity
Azure crossed $100 billion in annual revenue and grew 43% last quarter, yet Microsoft's CFO says cloud demand still exceeds the capacity available to serve it.
Xen 4.22 Adds Xenstore Quotas and Zen 5 Bus Lock Threshold
Xen 4.22 shipped on July 30, 2026 with per domain Xenstore quotas, AMD Zen 5 Bus Lock Threshold support on x86, PSCI suspend to RAM on Arm and FF-A 1.2.
YSERVER 1.4 Is a Rust X11 Server That Now Runs Steam
YSERVER 1.4.0 gives Chrome and Steam real GPU acceleration on a from scratch Rust X11 server, written largely with Claude Code and passing 672 XTS5 tests.
AMD Leases 530 MW of Data Center Capacity for 15 Years
AMD signed fifteen year leases with Core Scientific for about 530 MW of US capacity, worth over $14 billion, plus a reservation right on nearly 2 GW more.
AMD P-State Patches Lift 1% Low Frame Rates by 31.8%
A Meta kernel engineer proposes per core EPP boosting for amd_pstate. On a Steam Deck it lifts 1% low frame rates by 31.8% in Civilization VI.
ARCTIC Ships a Fan Controller That Linux Already Supports
ARCTIC launched a ten channel PWM fan controller whose hwmon driver was upstreamed months before the product. It lands in Linux 7.2 and works with lm-sensors.
Fish Audio Turns 31,000 GitHub Stars Into a $52M Seed
A voice model trained on one gaming GPU became Fish Speech, then $21M of ARR across 8 million users. Three of the five models stay open, one does not.
GCC Will Decline LLM Written Patches, With One Exception
The GCC Steering Committee adopted a policy declining legally significant contributions derived from LLM output. Test cases are carved out, and it is provisional.
Meta Keeps 20% of a 1 GW Campus and Leases Back the Rest
Meta and BlackRock will develop a $14B, 1 gigawatt campus in El Paso. BlackRock funds own 80%, a $12.5B debt package carries it, Meta leases the whole thing.
Starling: a Linux Desktop in Swift, Written With Claude
One developer, six months, 335,000 lines. Starling ships a Wayland compositor, an X11 server and a Swift shell, and runs Chrome and Slack unmodified.
Valve Pays Collabora to Port RADV to Windows
Collabora has RADV, Mesa's open source Vulkan driver for AMD GPUs, running Counter-Strike 2 on Windows. Valve funds it, and the kernel interface is the blocker.
FreeBSD Wrote Its Own NTSYNC Driver, Without Linux Code
The Q2 2026 status report runs to 41 entries. The headline is a clean room NTSYNC driver matching the Linux 7.0 interface, with an AMD ROCm port in progress.
MCP Goes Stateless Today, and Sticky Sessions Go With It
The 2026-07-28 revision of the Model Context Protocol drops the initialize handshake and Mcp-Session-Id, so any request can land on any server instance.
Moonshot Ships Kimi K3 Weights, All 1.4 TB of Them
Moonshot AI published the full weights of its 2.8 trillion parameter Kimi K3 on Hugging Face, along with the attention kernels and MoE communication library behind it.
NVIDIA Puts $5B Into Safe Superintelligence
NVIDIA takes a reported five billion dollar stake in Ilya Sutskever's lab and gives it Vera Rubin systems, moving SSI off Google TPUs and onto GPUs.
Roblox Opens Its Prompt to Game Alpha, Judged on Retention
Build enters public alpha in New Zealand on July 28. A phone turns a prompt into a playable game, then faces the same discovery ranking as every other Roblox title.
Ubuntu 26.04 LTS Gets an Opt In Virtualization HWE Stack
Canonical will ship qemu-hwe, libvirt-hwe, edk2-hwe and seabios-hwe on a six month cadence, so an LTS host can track newer QEMU without leaving the LTS.
GNU Binutils 2.47 Lands With a Faster Linker Setting
Binutils 2.47 is out with an ld -O 0 level that skips mergeable section merging, --start-lib for loose object files, and a batch of new RISC-V extensions.
Foxconn Drops VMware for a Two Year Old HCI Vendor
Foxconn replaced VMware and SAN storage across factories in China, Taiwan, Vietnam and North America with hyperconverged software from Arcfra, a company founded in 2024.
KVM Splits Its God Struct Ahead of Linux 7.3
Paolo Bonzini merged the kvm-chainsaw branch, breaking struct kvm_mmu into three. No new feature, but a large piece of nested virtualization confusion goes away.
Linux 7.2-rc5 Is Big, and a Third of It Is Networking
Linux 7.2-rc5 landed on July 26 with over 35% of its changes in networking drivers. Torvalds calls it too big but not scary. Stable targets August 16.
NVIDIA Weighs a $250B Backstop for OpenAI's Ohio Campus
NVIDIA is reportedly in talks to guarantee $250 billion so OpenAI can lease SoftBank's 10 GW Piketon campus. The structure matters more than the number.
Shopify Rebuilt Its Theme Layer With 93% Less Code
Shopify's next storefront theme drops JSON heavy configuration for HTML and Liquid, cutting code by 93 percent. The reason given is that AI agents read HTML better.
Claude Opus 5 Lands at Half the Price of Fable 5
Anthropic shipped Claude Opus 5 on July 24 at five dollars per million input tokens, half what Fable 5 costs, and it wins several published benchmarks.
One Line Fault Dropped 3 GW of Data Center Load in Virginia
A transmission fault in Ashburn sent about 3 GW of data centers to backup power in seconds. Nobody lost power, and that is exactly what worries grid operators.
Debian Votes Again on LLM Written Contributions
Four competing proposals, from an outright ban to conditional acceptance, went into discussion on July 24. Whatever Debian picks, derivatives will copy it.
glibc 2.44 Adds /etc/tunables.conf for System Wide Tuning
Released July 25, glibc 2.44 lets you set tunables in a config file instead of an environment variable, plus a THP switch and faster math.
Fifty Companies Sign an Open Weights Letter to Washington
NVIDIA, Microsoft, Meta and dozens more asked US policymakers not to restrict downloadable model weights. The signatory list doubled in a day and OpenAI joined.
Ubuntu Moves to Delete /etc/debian_version in 26.10
A twenty year old Launchpad bug is finally getting closed. If your scripts read /etc/debian_version to detect the distribution, now is the time to fix them.
Apple Opens MapKit for Automotive, Ford Ships It First
Apple announced MapKit for Automotive, an SDK that puts Apple Maps in the car as native navigation without CarPlay. Ford's 2027 EV platform is the first user.
FLUX 3 Puts Image, Video, Audio and Robotics in One Model
Black Forest Labs launched FLUX 3, one set of weights that generates images, 20 second video with native audio, and robot actions. Open weights come later in 2026.
Intel Data Center Revenue Grows 59% on Xeon 6 Demand
Intel posted $16.1B in Q2 2026 revenue, with data center and AI up 59% to $6.3B on Xeon 6. Demand now outpaces supply, which changes how you plan a refresh.
Seven Stable Linux Kernels Shipped in One Day
Greg Kroah-Hartman released seven stable Linux kernels on July 24, from 5.10.261 up to 7.1.5, one of them carrying over 2,000 patches. Here is how to stage it.
Qualcomm Raises Chip Prices by Double Digits in September
Qualcomm told customers on July 24 that chip prices rise by a double digit percentage for shipments after September 1, blaming supplier costs it can no longer absorb.
Stripe in Talks to Buy OpenRouter for About $10 Billion
Stripe is reportedly in talks to acquire OpenRouter for close to $10 billion, roughly eight times the model routing startup's $1.3 billion valuation from May.
AMD Launches MI400 GPUs and EPYC Venice at Advancing AI
At Advancing AI 2026 on July 23, AMD launched its Instinct MI400 GPUs, EPYC Venice CPUs, and the Helios rack, with Anthropic committing to up to 2 gigawatts.
AMD and Cerebras Split AI Inference Across Two Chips
AMD and Cerebras will pair Helios racks with the Wafer-Scale Engine for disaggregated inference, claiming up to 5x tokens per watt, a modelled figure, in late 2026.
Claude Voice Mode Adds Opus and Sonnet, Not Just Haiku
Anthropic updated Claude voice mode on July 23 so you can pick Opus, Sonnet, or Haiku, with voice actions in Gmail, Calendar, Slack, Canva, and Notion.
Etched Raises $300M at $10.3B to Scale Its Sohu Chip
Etched raised $300 million at a $10.3 billion valuation to scale Sohu, its transformer-only inference chip, in a round led by Sequoia with a16z and SK hynix.
Claude Cowork Turns Screen Recordings Into Reusable Skills
Anthropic shipped Record a Skill in Claude Cowork on July 21. Screen record a task once, narrate it, and Claude turns the demo into a reusable Skill.
Block Launches Buzz, an Open Source Slack and GitHub Rival
Block released Buzz on July 21, a free Apache 2.0 open source workspace built on Nostr where humans and AI agents share channels, code, and portable identity.
NVIDIA Spectrum-6: 102.4 Tbps Ethernet for AI Factories
NVIDIA launched Spectrum-6, a 102.4 Tbps Ethernet switch for the Vera Rubin platform that doubles capacity and holds 95 percent efficiency past 100,000 GPUs.
OpenAI Presence Puts Enterprise AI Agents Behind Guardrails
OpenAI launched Presence, a limited availability platform for deploying voice and chat AI agents with policies, guardrails, simulations, and Codex driven evaluation.
OpenAI's Project Camellia: a 3.2 GW Data Center in Georgia
OpenAI unveiled Project Camellia, a 3.2 gigawatt data center in Effingham County, Georgia, on a 25 year Georgia Power deal with up to 1,000 MW of demand response.
General Compute Raises $400M for an Inference Cloud
General Compute secured up to $400M from Upper90 to build an inference neocloud on SambaNova SN40 and SN50 chips, using inference silicon as loan collateral.
Google Gemini 3.6 Flash Lands in Copilot, 17% Fewer Tokens
Google released Gemini 3.6 Flash on July 21, a cheaper and more token efficient Flash model tuned for agentic coding, now rolling out in GitHub Copilot.
Kubernetes 1.37 Code Freeze: containerd 2.0 Now Required
Kubernetes v1.37 hit code freeze on July 22, 2026. It requires containerd 2.0, makes cgroup v2 mandatory and ships DRA and etcd RangeStream changes.
OpenAI Opens ChatGPT Work to Small Businesses
On July 21, 2026, OpenAI launched a ChatGPT for small businesses program built on ChatGPT Work, an agent that acts across apps and files with partner skills.
Alibaba Previews Qwen3.8 Max, a 2.4T Parameter Model
Alibaba previewed Qwen3.8 Max, a 2.4 trillion parameter multimodal model it ranks second only to Fable 5, with open weights promised soon.
China's CXMT Prices an 8.6 Billion Dollar DRAM Memory IPO
CXMT priced an 8.6 billion dollar Shanghai IPO, Asia's largest of 2026, as the world's fourth largest DRAM maker scales up server and AI memory production.
Hut 8 Fully Contracts Its 1 Gigawatt Beacon Point AI Campus
Hut 8 signed a second 15-year, 352 MW lease worth 9.8 billion dollars, fully commercializing its 1 gigawatt Beacon Point AI data center campus in Texas.
Microsoft Will Run AMD Helios Rack Systems on Azure
Microsoft will deploy AMD Helios rack-scale systems on Azure, pairing Instinct MI455X GPUs, EPYC Venice CPUs, and Pensando networking to serve frontier AI.
Z.AI Powers Up a 1 Gigawatt Data Center on Chinese Chips
Z.AI has begun operating a 1 gigawatt data center running only Chinese-made chips, training its GLM models on Huawei Ascend hardware and the MindSpore stack.
Cloudflare Puts Feature Flags in the Terminal With Wrangler
Cloudflare's Wrangler CLI now manages Flagship feature flags from the terminal, so you can create flags, flip kill switches, and roll back from CI or a script.
EU Orders Google to Open Android and Search to AI Rivals
The European Commission issued binding DMA measures on July 16, telling Google to open eleven Android features to rival AI assistants and share anonymized search data.
NotebookLM Is Now Gemini Notebook, With a Cloud Computer
Google renamed NotebookLM to Gemini Notebook on July 16 and gave every notebook a secure cloud computer that writes and runs code against your own sources.
SAP Buys Prior Labs to Lead Tabular Foundation Models
SAP has completed its acquisition of Prior Labs, the German lab behind TabPFN, backing a European frontier AI lab with more than one billion euros over four years.
Walden Robotics Exits Stealth With 300 Million for Robots
Walden Robotics launched from stealth with 300 million dollars at a 1.1 billion valuation, and its general purpose robots are already working on a Toyota line.
Apple Overtakes Nvidia as World's Most Valuable Company
For part of Friday July 17, 2026, Apple passed Nvidia as the world's most valuable company near 4.9 trillion dollars. What the crown swap means.
Equinix Tops Out Its MU4 Data Center Expansion in Munich
Equinix has topped out its MU4 expansion in Munich, adding room for thousands more cabinets at one of Germany's key interconnection metros.
Japan Builds World's First National AI Factory With Nvidia
Japan's FRONTia project puts 27,500 Nvidia Rubin GPUs in a 140 MW national AI factory to train robotics models. Inside the Noetra plan.
Spectro Cloud Raises 100 Million for Kubernetes AI
Spectro Cloud raised more than 100 million dollars in a Series D led by Goldman Sachs to help teams put AI on Kubernetes into production across clouds.
Anthropic in Talks to Lease 10 Billion in Compute From Meta
Anthropic is in early talks to lease as much as 10 billion dollars of compute from Meta over two years. What Meta Compute means for the cloud market.
Fireworks Raises 1.5 Billion Series D for Inference
Fireworks AI closed a 1.5 billion dollar Series D at a 17.5 billion valuation. Why specialized open model inference is the story developers should read.
Google Delays Gemini 3.5 Pro Over Coding and Reasoning Gaps
Google has delayed Gemini 3.5 Pro after the model missed internal targets on coding and long horizon reasoning, pushing the flagship months past its June date.
Thinking Machines Ships Inkling, a Leading US Open Model
Thinking Machines released Inkling, a 975 billion parameter open weight MoE with 41B active, a 1M token context, and weights on Hugging Face under an open license.
Moonshot Releases Kimi K3, the Largest Open Model Yet
Moonshot AI shipped Kimi K3, a 2.8 trillion parameter open MoE model with a 1M token context and thinking mode, with open weights due July 27.
PostgreSQL 19 Beta 2: REPACK, Graphs and New Defaults
PostgreSQL 19 Beta 2 shipped on July 16, 2026, bringing the REPACK command, SQL/PGQ property graphs, JIT off by default and lz4 TOAST compression.
Anthropic Lines Up Investor Meetings Ahead of an IPO
Bankers at Goldman Sachs, Morgan Stanley and JPMorgan are scheduling investor meetings for Anthropic, pointing to a listing as soon as October 2026.
Apple Intelligence Clears China With Alibaba's Qwen
China's Cyberspace Administration approved Apple Intelligence on July 15, 2026, on the back of a deal to run Alibaba's Qwen model inside iOS, iPadOS, macOS and visionOS.
Apple Opens Its New AI Siri to Everyone in the iOS 27 Beta
Apple released the iOS 27 public beta on July 13, 2026, putting the redesigned AI Siri in front of everyone. For developers, App Intents is now the only way in.
Emergent Raises 130M at 1.5B for AI App Building
Emergent, an Indian AI coding platform launched in June 2025, raised a 130 million dollar Series C at a 1.5 billion valuation with 120 million in run rate revenue.
etcd 3.7 Lands: Streamed Ranges and Lower Kubernetes CPU
etcd v3.7.0 shipped in early July 2026 with RangeStream, keys only queries and a lighter startup path, cutting etcd CPU use for Kubernetes clusters versus 3.6.
TYLsemi Raises 43M to Ship Custom AI Silicon via Chiplets
TYLsemi left stealth on July 14, 2026 with a 43 million dollar round and a full stack chiplet platform for custom AI silicon, with TSMC samples due in 2027.
ASML Lifts 2026 Outlook as Intel Ships High NA EUV Chips
ASML reported 9.3 billion euros of Q2 net sales and raised its 2026 outlook to 43 to 45 billion euros, as Intel became the first to ship logic made with High NA EUV.
Linux 7.2 Adds Cache Aware Scheduling for Big Server CPUs
Linux 7.2 reached rc3 on July 12, 2026, carrying cache aware scheduling, an Intel led feature that keeps tasks sharing data on the same last level cache.
New York Halts New Hyperscale Data Centers for a Year
Governor Hochul signed an executive order pausing state environmental permits for new data centers of 50 megawatts or more, for up to one year.
Rakuten Mobile Finishes Nationwide Open RAN Radio Rollout
Rakuten Mobile has deployed 1Finity Open RAN radios across its nationwide 5G network in Japan, with a further 3,000 massive MIMO units planned at 3.7 GHz.
Reflection Signs 1 Billion Dollar Nebius Compute Deal
Reflection AI signed a compute deal worth over 1 billion dollars with Nebius, running through 2029 and giving it access to Nvidia GB300 chips to train open models.
Tower Puts 3 Billion Dollars Into Japan for Optical Chips
Tower Semiconductor is investing about 3 billion dollars in Japan for 300mm silicon photonics and silicon germanium, backed by roughly 1 billion in grants.
Cloudflare Charges AI Agents Per Request via HTTP 402
Cloudflare's new Monetization Gateway uses the x402 protocol and HTTP 402 to let any site charge callers per request for pages, APIs, datasets and MCP tools.
Intel Puts 5 Billion Euros Into Ireland for More Xeon
Intel is investing 5 billion euros in its Leixlip campus in Ireland to scale Xeon 6 and next gen Xeon on the Intel 3 node, after buying back full control of Fab 34.
Meta Moves Its Iris AI Chip Into Production in September
An internal memo says Meta will put its fourth generation MTIA chip, Iris, into production in September 2026 to double compute and lean less on Nvidia and AMD.
Mistral Open Sources Leanstral 1.5 for Lean 4 Proofs
Mistral released Leanstral 1.5, an open weights Apache 2.0 model for Lean 4 that proves code correct and found 5 unknown bugs across 57 real repositories.
OpenAI Ships GPT-5.6 and ChatGPT Work, a Desktop Agent
OpenAI released the GPT-5.6 family (Sol, Terra, Luna) plus ChatGPT Work, a desktop agent with Codex built in that turns a goal into finished docs, sheets, and code.
Meta Bets 10 Billion Dollars on a 1 GW Alberta Data Center
Meta will spend about 10 billion dollars on its first Canadian data center in Alberta, drawing near 1 gigawatt. Why power, not chips, now sets the pace.
SambaNova Raises 1B at 11B to Challenge Nvidia on Inference
SambaNova closed the first part of a 1 billion dollar Series F at an 11 billion valuation to scale on premises AI inference. Why the anti Nvidia bet grew.
SK Hynix 26.5B Nasdaq Debut Bets Big on AI Memory
SK Hynix raised 26.5 billion dollars on Nasdaq, the biggest foreign US IPO ever, to fund HBM and fab expansion. What it means for AI memory supply.
How to Run Qwen 3.7 Locally with Ollama (2026)
How to run Qwen locally in 2026: Qwen 3.7 open weights are not out yet, so we set up the newest open Qwen with Ollama, by size, with a one-tag swap for 3.7.
The awk Cheatsheet: Process Text in Columns
A practical awk cheatsheet: print columns with $1 and $NF, set field separators with -F, filter rows by pattern, sum a column, and use BEGIN and END blocks.
dig and nslookup Cheatsheet: Query DNS Fast
A practical dig and nslookup cheatsheet: look up A, AAAA, MX, TXT, NS and CNAME records, trim noise with +short, trace from the root with +trace, and query @8.8.8.8.
Docker Commands Cheatsheet: The Everyday Ones
A practical Docker cheatsheet: run a container with docker run, list with ps, follow logs, exec into a shell, bring a stack up with compose, and prune disk.
The Git Cheatsheet: Everyday Commands and Quick Fixes
A practical git cheatsheet: check status, stage and commit, switch branches, and get out of trouble with reset, restore, revert, stash and reflog.
The sed Command Cheatsheet: Substitute and Delete
A practical sed cheatsheet: substitute with s/old/new/g, edit files in place with -i and a backup, delete lines, and print specific lines with -n p.
AI in Cybersecurity 2026: Real vs Hype, Both Sides
AI in cybersecurity in 2026: what it really does for defenders (triage, summarization, code review), what attackers do with it, and the dual-use problem.
Anthropic Suspends Claude Fable 5 and Mythos 5
Anthropic disabled Claude Fable 5 and Mythos 5 worldwide after a US government order. The timeline, the trigger, and how to fall back to Opus 4.8 fast.
Bug Bounty Starter Guide for Sysadmins
A bug bounty starter guide for sysadmins: how your DNS, HTTP and Linux skills transfer, the five free tools I open, and the report shape that gets paid.
Build a Local AI Agent with Ollama + Qwen 2.5
How I run a local AI agent with Ollama and Qwen 2.5 in 2026: install on Windows, Linux or macOS, pick the model your VRAM fits, call it from Python and LangChain.
CentOS to Rocky Linux 9 Migration Playbook 2026
CentOS to Rocky Linux 9 migration, end to end: the in-place ELevate route, the fresh-install route, the pre-flight inventory, and the breaks that fail silently.
ChatGPT vs Claude vs Gemini 2026
A hands-on ChatGPT vs Claude vs Gemini comparison for 2026, ten real jobs tested through the APIs: code, research, writing, data, vision, cost and latency.
The chmod and chown Cheatsheet, Decoded
A practical Linux chmod and chown cheatsheet: the Permission denied fix with chmod +x, why 755 and 644 keep showing up, chown for ownership, and recursing safely.
Choosing a Linux Distribution in 2026: Decision Guide
Choosing a Linux distribution in 2026 without the tribal noise: four inputs that decide (team skills, compliance, app stack, support) plus honest distro profiles.
Claude Fable 5.1 vs Opus 5: cost, caching and migration
Compare Fable 5.1 and Opus 5 using current API prices, a reproducible cost example and a migration checklist. Updated September 8, 2026.
Core Web Vitals 2026: LCP, INP and CLS Mastery
Master Core Web Vitals in 2026: what LCP, INP and CLS really measure, lab vs field data, the fixes that still hold up, plus budgets and a weekly workflow.
Build a CTF Lab on Hyper-V: Full Walkthrough
Build a CTF lab on Hyper-V: three trust-segmented virtual switches, a Kali attacker, Metasploitable3 and a broken Windows 10, with pfSense in the middle for routing.
curl Cheatsheet: Download Files and Test APIs
A practical curl cheatsheet: download and keep the remote name with curl -O, follow redirects with -L, fetch headers only with -I, and POST JSON with -d.
DNS over HTTPS Implementation Guide
DNS over HTTPS the way I roll it out: what DoH buys you (and what it does not), flipping it on for Windows, Linux and macOS, plus self-hosting and edge enforcement.
Fix: ESXi Host Not Responding in vCenter
ESXi host Not Responding in vCenter is almost always a dead management link, not a dead host. Restart hostd then vpxa, clear the log partition, check ports 902/443.
The find Command Cheatsheet: Recipes You Actually Use
A practical Linux find cheatsheet: find by -name, big files with -size +100M, recent ones with -mtime, delete with -delete, and run -exec on each hit.
Git Rebase vs Merge: A Visual Workflow Guide
Git rebase vs merge, decided in one mental model and five real situations: branch diagrams before and after, the exact commands, and the rule I never break.
The grep Cheatsheet: Search a File, Search a Tree
A practical Linux grep cheatsheet: search one file with grep pattern file, search a whole tree with grep -rn, ignore case with -i, invert with -v, and pick -E or -F.
Hyper-V: The Hypervisor Is Not Running, Fix Guide
Fix Hyper-V 'the hypervisor is not running' by working the five gates in order: BIOS virtualization, the feature, hypervisorlaunchtype, conflicts, services.
How I Optimize Images for Website Speed in 2026
Image optimization techniques for 2026: why AVIF goes first with WebP and JPEG behind it, responsive srcset, the lazy-load plus LCP trap, compression, and CDN automation.
INP Optimization 2026: Fix Interaction to Next Paint
A hands-on INP optimization guide for 2026: the twelve Interaction to Next Paint killers I keep finding in audits, how to profile them, and a real 480ms to 120ms fix.
iptables vs nftables: A Linux Firewall Cheatsheet
iptables vs nftables, side by side: list rules, allow SSH on dport 22, default drop policy, persist to rules.v4 or nftables.conf, convert with iptables-translate.
Linux Firewall Comparison 2026: ufw, nftables, iptables
Linux firewall comparison for 2026: iptables, nftables, firewalld and ufw all poke one kernel. Per-distro defaults, the same rule four ways, plus my honest pick.
Linux Kernel Hardening: sysctl Checklist
Linux kernel hardening with sysctl: the 32 toggles I actually ship, grouped by subsystem and sorted by how much each one matters, plus a drift check.
Linux Networking Commands: ip and ss Cheatsheet
ifconfig and netstat are gone on most modern Linux installs. The iproute2 replacements you need: ip a, ss -tulpn, ip route, ip neigh, and a clean old-to-new map.
MikroTik vs Ubiquiti for Home Networks (2026)
MikroTik vs Ubiquiti for a prosumer home network: matched stacks under 400 EUR, CLI vs UI, real routing protocols, Wi-Fi roaming, VPN, a 3-year cost breakdown.
Mobile Optimization 2026: Fast and Findable on Phones
Mobile optimization in 2026: the Core Web Vitals targets for mobile (LCP under 2.5s, INP under 200ms, CLS under 0.1), what slows a phone, and how to test like one.
Network Config Across Linux Distros in 2026
Network config across Linux distros in 2026: netplan, ifupdown, NetworkManager, systemd-networkd and wicked compared, with the static IP recipe for each.
nmcli Cheatsheet: Manage Wi-Fi From the Terminal
A practical nmcli cheatsheet for NetworkManager: scan and connect to Wi-Fi, list saved connections, bring them up and down, check device status, and set a static IP.
pfSense vs OPNsense for Home Networks (2026)
pfSense vs OPNsense for home networks, run side by side on one N100 mini-PC for years: UI, update cadence, plugins, real benchmarks, appliances, and which I pick.
PowerShell Networking Cheatsheet for Windows
A PowerShell networking cheatsheet: test a port with Test-NetConnection, read your IP and DNS with Get-NetIPConfiguration, find listening ports, resolve names.
Proxmox vs ESXi vs Hyper-V in 2026: Which to Run
Proxmox vs ESXi vs Hyper-V in 2026: honest cost, licensing, clustering, backup and containers after the Broadcom shakeup, plus which fits homelab or enterprise.
Python regex cheatsheet: 20 essential patterns explained
A Python regex cheatsheet of 20 tested re patterns: email, IPv4 and IPv6, URL, phone, ISO 8601, UUID, CSV split, log lines, slugs, with the edge case for each.
The rsync Cheatsheet: Mirror, Sync and Copy Over SSH
A practical Linux rsync cheatsheet: mirror with rsync -av, copy over ssh with -avz, mirror exactly with --delete, dry-run, resume, and the trailing slash gotcha.
Ubuntu 24.04 Server Hardening: SysAdmin Checklist
Ubuntu 24.04 server hardening, the way I run it on every VPS: keys-only SSH on a quiet port, UFW, fail2ban, unattended updates, AppArmor enforcing, auditd.
Self-Hosted VPN with WireGuard: Setup Guide
Self-hosted WireGuard VPN on a 5 EUR VPS: pick split vs full tunnel, generate Curve25519 keys, write the server and client configs, and get NAT and DNS right.
Nagios on Hyper-V: Server Monitoring Setup Guide
Server monitoring with Nagios on a Hyper-V Ubuntu 24.04 VM: build Nagios 4.5 from source, aim HTTP, SSH, MySQL and disk probes, then wire mail and Telegram alerts.
SOC Homelab with Wazuh, Suricata and ELK
Build a SOC homelab with Wazuh, Suricata and the Elastic Stack on a spare VM: the single-node installer, agent enrolment, and test attacks you can fire yourself.
SSH Command Cheatsheet: Connect, Keys, scp, Tunnels
A practical SSH command cheatsheet, most-used first: connect with ssh user@host, make an ed25519 key, push it with ssh-copy-id, copy files with scp, tunnel ports.
SSH Permission Denied (publickey): The Fix
Permission denied (publickey) is never mysterious once you read the handshake. The real causes in order, each with the exact command to confirm and fix it.
systemctl and journalctl Cheatsheet
A systemctl and journalctl cheatsheet: status to see if a service runs and why it died, restart, reload, enable --now, daemon-reload, and journalctl -u NAME -f.
Tailscale vs WireGuard in 2026: Not Competitors
Tailscale vs WireGuard in 2026: Tailscale is WireGuard plus a control plane for keys, NAT traversal via DERP, MagicDNS and ACLs. When each one wins.
The tar Command Cheatsheet: Create, Extract, List
A practical Linux tar cheatsheet: create a gzip archive with tar -czf, extract with tar -xzf (eXtract Ze File), list with tar -tzf, plus exclude and strip-components.
Core Web Vitals 2026: Fix LCP, INP and CLS (Real Code)
Core Web Vitals in 2026: before/after code I shipped for LCP, INP and CLS, reading a Lighthouse capture honestly, plus the WordPress and Cloudflare toggles I reach for.
Vaultwarden Self-Host Guide: Bitwarden in 20 Minutes
Vaultwarden self-host guide: docker compose plus Caddy with automatic TLS, your first admin signup, then fail2ban, restic backups, and locking down the admin route.
VMware ESXi PSOD: Diagnose and Recover the Host
VMware ESXi PSOD fix guide: read the purple screen, tell a bad driver from dying hardware, grab the coredump before you reboot, then recover the host.
VMware Workstation and Hyper-V on the Same PC (2026)
The VMware Device/Credential Guard not compatible error has two fixes: keep Hyper-V and upgrade VMware to Workstation 16+, or disable Hyper-V and VBS for VMware 15.
Wazuh SIEM on a Budget: Pi, N100 or Cloud
Wazuh SIEM on a budget, three paths I actually run: a Raspberry Pi 5 under 100 EUR, an Intel N100 mini-PC, or a Hetzner CAX21 ARM node under 10 EUR a month.
Web Accessibility and WCAG in 2026
Web accessibility in 2026: target WCAG 2.2 AA, fix the five issues that sink most sites (contrast, alt text, labels, link text, page language), and test it yourself.
Web App Security Audit 2026: A Practical Method
The web app security audit I actually run in 2026, in four layers: transport, app code, supply chain, secrets, with the tools and a 25-point checklist.
Web Workers in 2026: Patterns, Pitfalls and INP
Web Workers in 2026: the patterns I reach for, the ones I skip, transferables, SharedArrayBuffer, Comlink, and how to prove the move actually helped your INP.
WordPress Security Audit in 10 Steps (2026)
A WordPress security audit you run in about 90 minutes: PHP and WP versions, plugin CVEs, HTTP headers, htaccess, TLS, 2FA, xmlrpc, user enumeration, logs and backups.