tech

2025-05-05 · open the original ↗

• The GL.iNet Slate 7 was "absolutely perfect" as a travel router. On its first trip, at RSAC 2025, the host ran it as a wireless repeater because his hotel room had no Ethernet port, with his laptop hardwired to it through an Anker multiport hub; the kit also includes a YubiKey and a 1080p portable second monitor.
• The host's roughly four-year-old Samsung laptop needs replacing. DaVinci Resolve crashed several times on the trip while editing in only 1080p, and he is looking for a laptop for video editing.

2025-05-05 · open the original ↗

• AI development will reach many more non-developers. Dell's Logan Lawler says that, not being a developer when hired about a year earlier, he taught himself to pull SDKs from GitHub and train his own LoRAs for an animation studio. He expects that kind of access to become far easier and more common.
• AI will simplify small everyday tasks. Lawler's example is a RAG system over his 70-year-old mother's thousands of unsorted recipe cards that would return her pumpkin pie recipe on request. He notes someone would still have to set it up.
• Agents will coordinate with one another, then move into the physical world. Nvidia's Sama Bali, echoing Jensen Huang's GTC keynote (about five weeks before release, the host notes), describes factory agents passing word of a raw-material shortfall to the floor agent and then the transport agent so fewer trucks are booked. She points to Bay Area driverless cars visibly improving week by week as physical AI, and adds her personal hope for guardrails and regulation.

2025-05-04 · open the original ↗

• NIM microservices exist because models change every few months. Bali says Nvidia packages open, proprietary and its own models as containers with TensorRT-LLM-style inference optimisation, so developers can swap Llama 3 for 3.1 or 3.2 without rebuilding the pipeline; NIM stands for Nvidia inference microservice. Lawler adds that a NIM loads with one line of code and arrives already tuned to the GPU it runs on, so the only check is that the model fits in GPU memory, 96 GB on the Dell Pro Max machine discussed.
• Nvidia AI Enterprise layers training, guardrails and recipes on top. Bali lists NeMo for building, fine-tuning and guardrailing models, and AI Blueprints as reference workflows to which companies add their own data, which she calls their differentiation. She says NIMs can be prototyped free at build.nvidia.com and downloaded free through Nvidia's developer program.
• CUDA libraries accelerate data science without code changes. Bali says CUDA's parallelism can cut training from weeks to days. By her account, RAPIDS cuDF mimics pandas and Polars APIs to speed preprocessing by 100x on RTX 6000 GPUs with one added line, and cuML, announced at GTC, speeds scikit-learn tasks by up to 50x.

2025-05-03 · open the original ↗

• Data-science acceleration comes down to the GPU. Dell's Lawler says what matters is how much fits on the GPU and how fast it runs, with the rest of the Pro Max system in support. He says the T2 tower, launched on the recording day, grew in capacity for extra drives and cards without much bigger footprint, and runs new Intel processors at a sustained 250 W, which he calls unique in the industry.
• Scarce cloud and data-centre GPUs push experimentation onto the desk. Nvidia's Bali says fine-tuning and running models is now mainstream, even for product marketers like her and Lawler, while data-centre resources are scarce. She cites a data scientist who had to queue in a Slack channel for an instance, and argues developers need a local sandbox to keep up with daily new models and techniques.
• Doubled GPU memory is the new generation's biggest feature. For the RTX Pro Blackwell GPUs, Bali calls the jump from 48 GB to 96 GB per GPU, with up to four in one workstation, the biggest feature, and "a game changer" for running, fine-tuning and serving models locally.

2025-05-02 · open the original ↗

• The captions list the rack's contents but give no cost. The host walks the UniFi Pro Max rack at RSAC 2025: redundant UDM Pro Max gateways, two USW Pro Aggregation switches, three 48-port USW Pro Max switches, four UniFi Power Amps (he owns one and praises its audio), two NVR Pros, two UNAS Pros and a Power Distribution Pro. No prices are stated, despite the title.

2025-05-02 · open the original ↗

• Question Answering builds a no-code Q&A chatbot from a document. Giard presents it as the tool for letting users ask natural-language questions and get answers drawn from a list of questions and answers, such as an FAQ, without writing code. It is not on by default: it is one of the Language Service's optional "custom features", selected as custom question answering when the resource is created.
• The feature requires an Azure search service, created alongside it. The creation blade asks for the search service's region and tier; Giard takes the S pricing tier for the language service and the standard tier for search, both in East US. Deployment took about two minutes, and the resource group then held both the language service and an Azure AI search service.
• Configuration happens in Language Studio, under a project per resource. He opens language.azure.com, which redirects to language.cognitive.azure.com; with several language services deployed, he first picks the right one under the settings gear's resources tab. Creating a custom question answering project sets one language for all projects in the resource (English here) and a default answer for unmatched questions, which he changes from "no answer found" to an apology.
• Sources can be a URL, chit-chat or an uploaded file. Giard uploads a document of movie-quote questions and answers, each pair separated by a blank line, and the service parses it into pairs. In the editor he adds a pair of his own ("hello" answered with "hi, how can I help you?"), notes alternate phrasings can be attached to a question, and says follow-up prompts can be added.
• A test pane checks matching before a one-click deploy. A paraphrase ("you look short") still returned the answer stored for the stormtrooper question, and the alternate phrasing returned the same greeting. An "include short answer response" checkbox allows a short and a long answer, which he suggests pairing with a "more details" button; deployment took about 30 seconds and makes the knowledge base callable from a chatbot.

2025-05-02 · open the original ↗

• MCP addresses the integration problem holding back agentic AI. The host says LLMs have always struggled to reach data beyond their training, and custom per-source connectors were brittle and could not scale. Anthropic introduced MCP in November 2024, but only in the past couple of months, he says, has it taken off.
• Adoption has been fast because MCP is open and model-agnostic. He cites early adopters Block, Apollo, Replit and Sourcegraph, and over a thousand community-built MCP servers by February. Any model, Claude, GPT-4 or open source, can use it and anyone can build an integration without permission, positioning it as a "USB or HTTP of AI integration".
• Dynamic discovery lets agents find tools without hard-coding. The host explains that agents detect available MCP servers and their capabilities automatically, so a new server for, say, a CRM is usable immediately. Anthropic provides pre-built servers for Google Drive, Slack and databases. He says MCP complements rather than replaces custom APIs, plug-in systems, LangChain-style frameworks and RAG.
• MCP is the action layer, not a silver bullet. He places it as a standard integration layer for the action component of agents. Its open problems are managing many tool servers, getting models to use tools well, an evolving standard, security and monitoring, and overkill for simple apps. He relays Anthropic's plans for remote servers with OAuth, an official registry, standardised discovery endpoints, streaming and proactive server behaviour.

2025-05-02 · open the original ↗

• DotnetSnes compiles C# to SNES ROMs by transpiling .NET IL to C. Shapiro, a staff video architect at GameChanger, built DNTC, a transpiler that walks a dependency graph from a root function and turns the IL stack machine into C. Inspired by Nim compiling to C, the aim is to run C# anywhere a C compiler exists, without the .NET team building ahead-of-time compilation for every platform. The C output then goes through an existing SNES C toolchain to the console's 65816 CPU.
• Hardware without dynamic memory allocation suits the transpiler for now. It does not yet support reference types because there is no garbage collector, which he named as his next task. That fitted eBPF, which he said Windows is also moving toward after the CrowdStrike outage, and he has proof-of-concept C# eBPF programs running in the Linux kernel. It also fitted the SNES, which has no dynamic allocation either.
• Attributes let C# and native C interoperate at zero cost. Calls work in both directions. Attributes for custom function names and native functions and globals map C# onto the C SDK's expectations, such as a lowercase main and assembly labels at fixed memory locations. Because everything links into one binary, native calls carry no overhead, which Hanselman compared to P/Invoke.
• The project stands on community libraries. It builds on Mono.Cecil for reading IL and on PVSnesLib, a C SDK for the SNES that grew from an older one. Because PVSnesLib targets non-C99-compliant C, the transpiler takes plugins to adjust its output. The SDK's support for Tiled maps gives collision detection, and the Mario sample is a near line-by-line port while the Breakout sample is more idiomatic C#.
• It runs easily today and has already run on real hardware. Hanselman reported building the Mario ROM in WSL in about 10 minutes, after updating submodules recursively and using SSH keys. Shapiro said someone ran the Breakout game on a real Super Nintendo from a flash cart. His next aim is a basic PlayStation 1 C# engine, after a Pikuma course.

2025-05-01 · open the original ↗

• The Language Service can be created without the full AI services set. Giard creates it alone in the Azure portal for when only a subset of Azure AI services (formerly Cognitive Services) is needed; sentiment analysis is among its standard features. He skips the optional custom features, such as custom question answering and custom text classification, because they add resources like a storage account and Azure AI search.
• Setup choices are a Free or S tier plus optional tabs. He takes the S tier, in East US, for more calls and higher capacity, and leaves the optional tabs at their defaults: network restrictions, the identity the service runs under, and tags (name-value pairs for sorting or filtering in reports). Creation took about a minute.
• Two keys exist so one can be regenerated without downtime. The endpoint and keys, under Manage keys, are what Giard says matter most. Only one key is needed; if one is compromised, callers switch to the other while it is regenerated.
• Sentiment analysis is a plain REST POST with a JSON body. He runs curl in a bash Cloud Shell, posting to the service's endpoint plus text/analytics/v3.0/sentiment, with the subscription key and a content type of application/json in the header. The body is a "documents" array; his one document carries a language (English), an id of 1 and the text.
• The response gives an overall label and three confidence scores. "Your product is great" came back positive with scores of 1.0 positive, 0.0 neutral and 0.0 negative, and "your product is terrible" negative at 0, 0 and 1.0. A mixed sentence ("not too bad… has some OK features") came back positive but at 0.92 positive, 0.07 neutral and 0.01 negative, which Giard reads as somewhere in between.

2025-05-01 · open the original ↗

• Dell collapsed its many PC brands into three tiers. Lawler says the rebrand to Dell, Dell Pro and Dell Pro Max replaces many disparate lines (Precision, OptiPlex, XPS and others) to match how customers shop. Dell Pro Max targets heavy GPU workloads from independent software vendors, such as engineering and Adobe workflows, and was announced at CES and launched at GTC.
• Purpose-built systems bring server power to the desk side. Lawler says the traditional Blackwell-accelerated towers and mobiles remain, but the new part is two system-on-a-chip Grace Blackwell machines purpose-built for data scientists and developers, which he presents as Dell recognising where the market is going.
• Nvidia's software layer makes the hardware a full solution. Bali, who leads Nvidia's AI solutions go-to-market, says its AI software gives a consistent, pre-optimised experience from workstation to server to data centre and back. Developers then need not tune models to hardware themselves; Nvidia and Dell sell the combination as the Dell AI Factory with Nvidia.

2025-04-30 · open the original ↗

• Ubiquiti showed two reference racks at RSAC 2025. The Pro Max rack ran redundant UDM Pro Max gateways, a USW Pro Aggregation, 48-port Pro Max switches, the UniFi power amp, NVR Pro and UNAS Pro units and a power distribution Pro; the Enterprise rack ran redundant Enterprise Fortress Gateways, an Enterprise 100G aggregation switch, four Enterprise 48s, three AI Keys and four Enterprise NVRs.
• The AI PTZ camera was on display ahead of release. The host says it is not yet released but "coming soon", shown in white at the main booth and black at a second booth; he likens it to the earlier large G4 PTZ, and separately says his own G5 PTZ tracks his dogs around the yard.
• The U7 Pro Outdoor turns directional on 6 GHz only. The host, seeing it for the first time, says it shares the U7 Outdoor's form factor but feels heavier, and with its antennas fitted it is directional on 6 GHz while staying omnidirectional on 2.4 GHz and 5 GHz.
• The Access demo paired the intercom with a separate intercom viewer. The viewer pulled the intercom's camera feed and audio live, with QR, directory, PIN and NFC unlock on the intercom; Ubiquiti's self-contained Access display case, built around a Cloud Gateway Max, is not a product it sells.

2025-04-30 · open the original ↗

• The one-shot fractal shader was partly luck. Hanselman puts the hand-written version at four to six hours, against under 15 minutes vibe coding; Mark says he gave it about an 80% chance with the example file supplied. The first attempt on Claude 3.7 thinking failed, and Mark argues that because output is non-deterministic, the success might have been a 10% draw, which is why some vibe coders run a prompt several times and pick the best result.
• Bigger context windows don't fix work on large codebases. Mark cites needle-in-a-haystack tests in Microsoft's Eureka benchmarks, where frontier models fail to connect two related facts placed far apart at 32k or even 8k tokens, far below the million- or 128,000-token limits. He also cites the "Lost in the Middle" paper, which found that models tend to ignore information in the middle of a long context.
• Vibe-coded v1 becomes an inherited codebase you must maintain. Mark says models struggle with brownfield changes: they lose track of state, forget to update functions, drop unrelated functionality while refactoring, and claim fixes they did not make. Hanselman calls the result "a crappy code base by a mediocre programmer", and at that point you may still have to spend the four to six hours learning HLSL.
• Enterprises shouldn't vibe code, because humans stay accountable. Mark's examples are unvetted packages that may carry vulnerabilities or bloat, models defaulting to a deprecated version of the OpenAI API because of training cutoffs, and specs that are only about 80% written up front. He concludes that, given transformer context limits, AI will not take over and experts must own the code, though agents that can check their own output will help.
• Code citations, licensing and indemnity matter in practice. Hanselman credits GitHub Copilot with flagging the Stack Overflow and GitHub sources it drew on, along with their licenses, which let them credit the authors in a README; Mark adds that Microsoft indemnifies Copilot users in copyright disputes. Separately, Hanselman dismisses Google's "25% of our code is AI generated" figure, since lines of code is "a lie" and the number may count autocompletes or unaccepted code.

2025-04-30 · open the original ↗

• GTC has become a festival-scale event spilling across San Jose. Bali, of Nvidia, describes partners from cloud providers to startups filling the town, many conversations with robots over five days, and people lining up from 6:00 a.m. to get into the keynote, which Jensen Huang preceded by serving food from a Denny's food truck.
• Attendees should arrive with a plan or be overwhelmed. Lawler, at his first GTC, says the volume of learning tracks and booths was overwhelming in a good way, and advises picking a learning track suited to your role beforehand rather than wandering the show floor.

2025-04-29 · open the original ↗

• Decide monolith versus microservices at deployment, not during development. Van Vliet's talk "Architectural Uncertainty: Build Schrödinger's App" argues that an application is in every state at once while it is being built, and the deployment moment is the "measurement" that settles it: one monolith, a few services or many, chosen by configuration.
• This only works if the code is a properly modular monolith. He treats modules as non-distributed microservices, drawn on the same boundaries and loosely coupled so they can be split off and spread across a network; tight coupling breaks it. He uses feature-based vertical slices, but says each slice is still layered (presentation, business logic, persistence), so it can be split horizontally too, such as gathering every feature's presentation components onto one server to act as a backend-for-frontend.
• His team's Jitar splits an application at runtime, from configuration alone. Van Vliet says Jitar ("just-in-time architecture") is an open-source distributed JavaScript runtime, deliberately not a framework, because a framework would mean rebuilding the application around it. It follows JavaScript's explicit imports and exports, places modules on the servers the configuration names, and turns a call between modules on different machines into a remote procedure call. Each configured segment becomes its own bundle, which he says is mostly built into a Docker image and deployed on Kubernetes, with more instances added for a segment under load.
• The main payoff he reports is simpler, safer development. Instead of spinning up mocks or every dependent service to test one microservice, developers run the whole application with one command, and modules call each other directly with no service contracts to keep up to date, so a change by one team shows up for the others at once.
• Jitar is experimental and has not reached version 1.0. Van Vliet says the team is still working out its value proposition and gathering feedback from conference talks, and had no clear answer yet on plugging it into an Azure DevOps pipeline. He points to jitar.dev and to Comify, the team's demo comic-based social media platform.

2025-04-29 · open the original ↗

• Scarce cloud compute is pushing AI development back to the desk. Bali, who leads AI solutions go-to-market at Nvidia, recalls a data scientist who had to queue in a Slack channel for an instance, and argues experimentation needs a local sandbox. Nvidia's RTX Pro Blackwell GPUs double memory from 48 to 96 GB per GPU, with up to four in one workstation.
• Nvidia's software layer is meant to remove hand-tuning models to hardware. Bali describes NVIDIA AI Enterprise: NIM microservices package models in containers already optimised for the GPU they land on, so a model can be swapped in with one line of code as Llama versions turn over every couple of months. She cites build.nvidia.com for free prototyping, NeMo for fine-tuning and guardrails, and CUDA libraries such as RAPIDS cuDF (up to 100x on pandas-style preprocessing, no code change) and cuML, announced at GTC (up to 50x on scikit-learn tasks).
• Data-science-only users should choose Dell's Linux GB10 and GB300. Lawler says these Grace Blackwell Dell Pro Max systems ship Linux-only with the Nvidia stack preloaded, while anyone also needing tools like Adobe Creative Cloud, which doesn't run on Linux, should take a traditional Windows Pro Max. Bali adds that the palm-sized GB10 can sit beside a Windows laptop, and two can be daisy-chained over ConnectX.
• The GB300 offers server-class power without server infrastructure. Lawler gives it 784 GB of unified memory (288 GB GPU, 496 GB CPU) and 20,000 TOPS at FP4, against roughly 4,000 for a single 96 GB RTX Pro 6000, and pitches it at mid-market firms with no racks or cooling. Pricing was unreleased at recording; he expects the GB10 near Nvidia's DGX Spark ("3 to 4,000") and the GB300 to cost more but less than a server.
• Both guests forecast AI spreading to non-developers and the physical world. Lawler, not a developer, says he trained his own LoRAs for an animation studio and imagines a RAG system over his mother's recipe cards. Bali, echoing Jensen Huang's GTC keynote, predicts a shift from generative to agentic AI, with agents handing off across a factory, and then physical AI such as driverless cars, adding her personal hope for guardrails and regulation.

2025-04-28 · open the original ↗

• The Pro XG24 PoE is a layer 3 multi-gig PoE switch. It has 8 2.5 GbE and 16 10 GbE RJ45 ports, all PoE++, plus two 25 Gbps SFP28 cages; as of recording, the host says the only UniFi gateway with SFP28 cages to match is the Enterprise Fortress Gateway (EFG).
• Hot-swappable power and fans mean stepping up to Enterprise Campus. The Pro XG24's fans and power are built into the chassis, with a locking power port and an RPS port for redundancy; hot-swap supplies or fan modules need the Enterprise Campus 24 PoE, about $700 more.
• The Pro XG family spans four models, including a desktop one. The Pro XG10 PoE has 10 10 GbE PoE++ ports and two 10G SFP+; the Pro XG48 PoE has 16 2.5 GbE and 32 10 GbE PoE++ ports and four SFP28; the $499 Pro XG8 PoE is, as far as the host knows, the first desktop switch with Etherlighting.
• Ubiquiti uses its own name for 90 W PoE. The host explains 802.3bt Type 3, known as PoE++, supplies up to 51 W, while Type 4 at up to 90 W carries an extra-plus name he believes only Ubiquiti uses; the only device he knows that accepts it is the Switch Flex 2.5G PoE.
• These are among UniFi's cheapest switches for powering 10G access points. The host names the Enterprise E7, U7 Pro and U7 Pro XGS as targets; the Pro XG24 offers 720 W of PoE and 460 Gbps switching capacity, drawing 150 W bare and up to 870 W fully loaded.

2025-04-28 · open the original ↗

• Large language models are here to stay, but every layer around them is in flux. Webber says integration, fine-tuning, agentic systems, pre-training, data-set selection and evaluation will all keep changing, while the core problem of training a model as well as possible and serving it efficiently has not changed and, she expects, never will.
• Knowledge belongs inside the model wherever possible. Drawing on her SageMaker years, Webber argues that pushing knowledge into the network simplifies development, data management and application management, and calls RAG and agentic layers "fluff" next to what the weights hold. She predicts problems first solved at the agentic level will be absorbed into RAG, then pre-training, then the data set itself.
• AWS's accelerator roadmap is only beginning, by her account. Webber says Trainium 3, pre-announced at re:Invent, is on the way, and that customers are "very much just getting started" with what Trainium and Inferentia will deliver.

2025-04-27 · open the original ↗

• Trainium and Inferentia share a core and differ in topology. Webber says both use the same NeuronCore and software stack, so workloads can move between them. Trn1 connects cards in a 4D torus suited to gathering results across cards for the backward pass and optimizer update, while Inferentia lays cards out more simply for sharding a large tensor through a forward pass.
• Inferentia offers more sizes because inference needs vary more. Webber says Inferentia instances come in many combinations of accelerator count and HBM capacity, suited to hosting a 7 billion or 11 billion parameter model, while Trainium comes essentially small and very large: develop on a single Trn1, then scale to as many large instances as you can get.

2025-04-26 · open the original ↗

• AWS builds its own accelerators to give customers hardware choice. Webber frames AWS's position as choice over data sets, models and accelerated hardware, and says customers pick Trainium and Inferentia above all for price-performance and energy efficiency. She claims Trn2 is the most powerful EC2 instance on AWS for AI/ML.
• Annapurna Labs' track record underpins the chips, in her telling. Webber says Amazon acquired the startup in 2015 to rebuild the hypervisor, and its Nitro system physically separates customer data from instance control, which she credits for both AWS's scaling and its security; every modern EC2 instance runs on it.
• Graviton CPUs now take most new AWS compute. Webber cites re:Invent: more than half of new compute coming onto AWS is Annapurna's Arm-based Graviton, flagged by a trailing "g" in instance names such as C6g. Trainium and Inferentia are its third product line, focused on AI/ML.

2025-04-25 · open the original ↗

• ASML alone can build the machines behind leading-edge AI chips. The host says the Dutch firm's latest lithography system weighs 150 tons, is roughly the size of two shipping containers and costs $350 million. No other company can reliably print features below 7 nanometres, and ASML holds over 90% of the market even for mature technologies, supplying TSMC, Samsung and Intel.
• EUV light comes from tin plasma about 40 times hotter than the sun. The host describes the machine firing 50,000 droplets of molten tin into a vacuum, each hit twice by lasers, producing plasma near 220,000°C. The light is bounced off mirrors with imperfections measured in trillionths of a metre onto a chip blueprint. The resulting processors can hold over 100 billion transistors, more than 70 layers and over 100 km of wiring on silicon about 1.5 times a US postage stamp.
• Lithography leadership is measured in decades, not months. The host argues that ASML spent two decades perfecting EUV and money cannot compress that. Its high-NA EUV raises aperture from 0.33 to 0.55 to print 8 nm features, with mirrors doubling in size and growing 10 times heavier. Hyper-NA above 0.75 is next, and wavelengths around 6 nm are seen as a plan B.
• China is squeezing older tools while building its own. Barred by the US from ASML's most advanced equipment, Chinese makers use multi-patterning on older machines to print features two or four times smaller, at a cost in complexity and speed. A state-owned firm is reportedly progressing on a 28 nm-capable tool. The host says matching EUV would mean replicating ASML's supply chain of more than 5,000 specialised suppliers.
• Canon's nanoimprint bet is cheaper but not yet chip-grade. Per the host, Canon's press-like nanoimprint lithography could in theory reach nanometre accuracy at about 40% lower cost. Defects, alignment and speed have so far confined it to uses like smartphone displays and memory chips, where higher defect rates are tolerable.

2025-04-24 · open the original ↗

• Never-ending support is a drop-in package feed, not a migration. Barnes, open source partner manager at HeroDevs, says customers get a token from the customer portal and point their package sources (package.json, Maven, Gradle) at the HeroDevs repository. From then on, high and critical CVEs in end-of-life versions are patched and pulled in on normal builds under the company's SLA. He describes HeroDevs as about 120 employees and mostly bootstrapped, helped by Google effectively blessing it as AngularJS's maintainer when that was deprecated.
• Paid support for old versions funds and relieves upstream maintainers. Barnes says HeroDevs hires key project staff, sponsors bodies such as OpenJS, funds community efforts and shares revenue upstream. In his account, projects can then point to HeroDevs in their deprecation notices and keep innovating on new versions. Hanselman frames it as a balance between open source and capitalism, though he says he will keep urging people to upgrade.
• Running unpatched software carries real regulatory and legal exposure. Barnes, a former practicing attorney, cites CISA rules requiring government contractors to patch vulnerabilities within set times, extra requirements in insurance, financial services and healthcare, and EU regulation. He adds that a breach traced to an unpatched, remotely exploitable high-severity CVE helps show negligence in a class action.
• Demand for end-of-life versions grows the longer they're dead. Barnes calls the growth "exponential", because companies and agencies rolling out mandatory CVE scanning keep finding forgotten apps; large .NET customers run tens to hundreds of applications. The current .NET focus is LTS releases that have gone end of life, starting with .NET 6, which ended support in November 2024. An STS version would be considered on demand, since the catalog is customer-request driven.
• Long-tail libraries are triaged per customer. At onboarding, HeroDevs reviews each dependency and decides whether to fork it internally or pay its maintainer to keep a branch, drawing on research into the support life cycles of the top 200 community NuGet packages. Builds must match distribution packaging bit for bit for apt and RPM repos, with a WinGet repo planned.

2025-04-24 · open the original ↗

• Size a FortiGate by its threat-protection throughput, not its ports. UTM inspection is CPU- and RAM-heavy: the FortiGate 70F lists over 1 Gbps of IPS and NGFW throughput but only 800 Mbps of threat protection. The host, from a firm that says it is a Fortinet partner, chose the 70F as about the smallest model supporting explicit proxy and SSL decryption.
• Several features are hidden from the menu by default. The 70F ships at 192.168.1.99 on port 1 with user admin and no password; options missing from a guide are under System > Feature Visibility. The host moves the LAN to 10.10.0.1 and adds employee (VLAN 20) and guest VLANs, tagged down to a UniFi switch as a third-party gateway network.
• A WAN cannot join an SD-WAN group while anything references it. The host deletes the factory LAN-to-WAN policy first, and configures WAN 2 even if absent so failover is plug-in later. His SLA pings Cloudflare and Quad9 every 500 ms, marking a link down after five failures and up after five successes; the rule prefers WAN 1 over WAN 2.
• Security profiles only act on the firewall policies they are enabled on. FortiGate denies everything not explicitly allowed, including inter-VLAN traffic; LAN-to-WAN rules need NAT, LAN-to-LAN rules should not have it. The host writes one rule per network so each can carry its own antivirus, web and DNS filter profiles, such as separate student and teacher filters.
• SSL inspection needs the FortiGate's CA certificate deployed to clients. Without it, every HTTPS page shows a certificate warning; pushed through something like Microsoft Intune, it is invisible. Without decryption, the host says, filtering still works but is less effective, since malware over HTTPS cannot be seen; the default antivirus profile blocks rather than monitors.

2025-04-24 · open the original ↗

• NxD hides most of the Neuron SDK's complexity from developers. Webber says the Neuron SDK spans low-level runtime, driver and compiler up to NeuronX Distributed (NxD), the main modelling library. NxD compiles a model and shards a checkpoint such as Llama or PixArt across an instance's accelerators, handling communication, optimizer updates and the forward pass. Prebuilt NxD model packages let users run alignment, supervised fine-tuning or hosting experiments with little exposed complexity.

2025-04-23 · open the original ↗

• A wrongly blocked domain is fixed by allowlisting or pausing Pi-hole. The host meets this most during online shopping; a temporary disable resumes blocking automatically when the timeout expires, but normally needs the Pi-hole web interface.
• Pi-hole Remote's paid version controls blocking from an Apple Watch. The free iPhone app connects to multiple Pi-hole instances and can disable blocking; the paid version, which cost the host $9.99 for a lifetime licence, adds the watch, with pauses from 1 minute to 8 hours, for the rest of the day, or indefinitely.

2025-04-23 · open the original ↗

• A kernel is where the user overrides the compiler on the chip. Webber explains that PyTorch code for Trainium and Inferentia passes through PyTorch XLA into an HLO graph, which AWS's compiler turns into the chip's instruction set. A kernel replaces that generated code with operations the user defines through NKI, the Neuron Kernel Interface.
• Kernel work is the study of compute optimisation for a workload. Kernels range from a hello-world tensor add to a full forward or backward pass or just an MLP projection, Webber says, and writing one means studying data movement, memory and compute utilisation to improve performance on large language models.

2025-04-22 · open the original ↗

• Customers pick AWS's own accelerators mainly for price-performance. Webber, who moved from SageMaker to the Trainium and Inferentia team because she saw accelerator supply and health become "make or break", frames the chips as part of AWS offering hardware choice. She traces them to Annapurna Labs, acquired in 2015, whose Nitro system and Arm-based Graviton CPUs (over half of new AWS compute, per re:Invent) preceded Trainium, and claims Trn2 is AWS's most powerful AI/ML instance.
• The Neuron software path lets PyTorch code reach the chip, with kernels as an override. Webber explains that PyTorch XLA turns a model into an HLO graph that AWS's compiler lowers to the chip's instruction set, while a kernel written in NKI, the Neuron Kernel Interface, replaces compiler output with user-defined operations tuned for data movement and utilisation. NeuronX Distributed (NxD) handles compiling and sharding checkpoints across accelerators, and Trainium also runs under SageMaker Studio and JumpStart.
• Trn2 quadruples Trn1's compute with the same NeuronCore. Webber gives Trn2 eight NeuronCores and four HBM banks per card versus two and two, 96 GB of HBM per card, 16 cards and 1.5 TB per instance. An UltraServer joins four Trn2 instances, 64 cards over two racks, via NeuronLink with at most two hops card to card. Because the core is unchanged, kernels and NxD code move up easily; she recommends Trn1 for models around 70 billion parameters and Trn2 for much larger ones.
• Tensor-parallel degree is a tuning knob with real costs. Webber corrects the host, who took TP for numeric precision: it is the number of NeuronCores sharing one copy of a tensor. Changing it, from TP8 on Trn1 to TP32–128 on Trn2, shifts time spent in collectives such as all-reduce, batch size and step time, and can help or hurt. Trn2's logical NeuronCore setting regroups 128 visible cores into 64 to give each more HBM bandwidth.
• Anthropic is the flagship customer, and academics get $110 million in credits. Webber cites Project Rainier, a very large Trainium cluster built with Anthropic, plus small-language-model startups and Databricks; the host adds Poolside. Build on Trainium offers $110 million in credits to academic researchers, scaling projects on a cluster of up to 40,000 Trn1 cards, and recently paid $25,000 for the fastest Llama implementation in NKI.

2025-04-21 · open the original ↗

• A model without web access may silently invent what a link says. The host recounts giving OpenAI's o1 pro a URL in deep research, at a time it lacked internet access; its reasoning trace said it would assume what the page contained, and he says the output alone would have misled him had he not checked the trace or the site.
• LLMs graded each other erratically but showed no self-preference. Michaelson describes generating hundreds of SAT essay questions, having four LLMs answer them, then having the same four rank the answers. Results were "wildly unstable" between reruns, but he saw no sign of a model favouring its own answers.
• LLM committees are his answer to evaluation bias. Michaelson still finds LLMs effective at judging LLM output, and suggests playing models off one another, for instance letting a cheap model screen answers and calling an expensive one only when an answer looks questionable. He mentions teacher-student distillation from a recent AWS talk in Ireland but says he hasn't looked into its specifics.

2025-04-20 · open the original ↗

• This upload holds a different clip than its title names. The video posted as "The LLM Revolution Threatens SaaS Businesses" is a copy of the 2025-04-19 clip on deploying models: both run 5:56 and their captions are identical. Greg Michaelson's SaaS argument is summarized under the 2025-04-17 clip and episode 879, dated 2025-04-15.

2025-04-19 · open the original ↗

• Handing models to engineers is a risky dependency, not just a bottleneck. Michaelson argues that the developers who recode data scientists' prototypes usually aren't data scientists and can easily introduce mistakes, while data scientists often lack the skills or infrastructure privileges to serialise, dockerise and deploy models themselves.
• Zerve packages environments and models so data scientists can deploy. Michaelson says every Zerve canvas runs in its own reusable, shareable Docker container, ending dependency mismatches for new hires, and every executed block serialises its variables. An API route can point directly at a trained random forest, and deployments run on Lambda-style serverless infrastructure with no long-running services. Where organisations still forbid data scientists from deploying, a handoff system passes deployable software rather than prototypes to other teams.

2025-04-18 · open the original ↗

• The NPU stayed idle until a Windows Insider build was installed. After that, the host saw it working on live captions for a YouTube video and in AMD's Amuse image generator running the SDXL Turbo Stable Diffusion model with Ryzen AI support.
• The NPU's gain for local AI is efficiency, not speed. The host says it does not necessarily run LLMs faster than a GPU but draws less power and heat; he expects real NPU support in tools like Ollama within a few updates, though for now it takes manual setup.
• The host rates it his likely favourite mini PC as a workstation. Testing since early February, he saw no slowdowns from the 12-core Ryzen AI 9 HX 370 with Radeon 890M graphics, but not for competitive gaming. It has dual 2.5 GbE, Wi-Fi 7, OCuLink and an internal power supply, from $929 with 32 GB and 1 TB, up to 96 GB and 2 TB.

2025-04-18 · open the original ↗

• DeepSeek set off an AI mania in Chinese markets and enterprises. The host reports that the Manus agent's registration site crashed within hours of its March 6 launch, with codes reportedly resold on the black market. The Hang Seng Tech Index rose more than 40% from mid-January, and hundreds of large firms, Tencent, city governments, hospitals and universities announced plans to integrate DeepSeek's models.
• Chinese investors are funding startups while calling it a bubble. One Hangzhou-based investor, quoted by the host, said there was "no choice" given the weak economy, with a strategy of entering at an A round and exiting at an A+ round a few months later. On March 6 the central government announced a venture fund of a trillion yuan, about $140 billion, for tech investment.
• Big tech is betting cloud AI will offset weak core businesses. Per the host, Alibaba declared human-like AGI its main objective, released a reasoning model it claims matches DeepSeek, and committed around $53 billion over three years to data centres, more than in the past decade, from a 36% share of China's cloud market. Baidu's cloud revenue has risen.
• The boom is driving a server and hardware spending surge. Demand for AI servers has soared since early February, including all-in-one on-premises units favoured by state-owned firms. A server maker founded by former Huawei staff is up about 140% in 2025, and Jefferies analysts estimate server makers may spend over 1.4 trillion yuan, about $200 billion, over two years.
• Chip access is the boom's chief vulnerability. Morningstar's Kai Wang argues DeepSeek won't fundamentally change most rally beneficiaries, and the host notes an earlier rally faded. Chinese firms still buy Nvidia's weaker H20, local designers such as Cambricon and Huawei lag, and SMIC is capacity-constrained. The host warns that reported US plans to restrict even the H20 could push costs back up and end the euphoria, while welcoming the open-sourcing DeepSeek leads.

2025-04-18 · open the original ↗

• Trust rests mostly on faith, but not only on faith. Synadinos says most definitions of trust involve faith or belief. Between blind faith and hard evidence he puts familiarity (an unbroken history of earlier trust) and authority (a boss or someone held in high regard).
• Risk decides whether trust is enough. Where the outcome could be catastrophic, or something bad is very likely, he says trust may be insufficient and actions should rest on more than faith. Where even the worst case is a "nothing burger", or very unlikely, trust is fine. The host offers his own pair of examples: lending someone a dollar against running an unknown executable that could compromise his network.
• Gathering evidence is costly and sometimes impossible. It takes effort, time and maybe money. Synadinos's example is bridges and elevators: failure could kill you, but you cannot interview their engineers, so you rely on faith. Trying to verify everything would mean analysis paralysis, so he calls trust a heuristic that lets the brain reach sensible conclusions without enough evidence.
• Skepticism is also wrong when misapplied. His flowchart, which he says is the slide audiences like best: pressing questions on a low-risk matter reads as cynical or jaded, while failing to question a high-risk one, such as plugging in an unknown flash drive, leaves you naive and easily manipulated. He says he gives people tools for striking the balance case by case rather than an answer.
• He built the "Trust but Verify" talk around misinformation and AI. He says he is despondent at seeing family and friends fooled, over-trusting some sources and under-trusting others. He moved the topic into a technology conference keynote by applying it to AI, which yields a mix of true and false information. Giard, who works for Microsoft, says the "Copilot" name is apt because it keeps the human as pilot. Synadinos closes his talk with a line from Wes Moore: "I want skepticism to be your companion and not your captor."

2025-04-17 · open the original ↗

• The book aims to keep middle schoolers by casting them as builders. Looper's Computer Science for Kids: A Storytelling Approach (Wiley) targets US grades 6 through 8, ages 11 to 13, the point where she says her research shows girls in particular drop out of computer science. Readers follow a guide through a mystical forest solving problems such as securing a gate against a moss thief, with a hidden curriculum on sustainability.
• Every chapter is mapped to a formal curriculum standard. The standards are the Computer Science Teachers Association's. Looper says they are demanding, down to compliance for sixth to eighth graders and training machine learning models from scratch by high school, and expects them to be revised toward AI this year and next. Hanselman singles out the mapping as what separates the book from whimsical kids' titles, along with its recurring "think like a computer scientist" sections on problem solving.
• Everything runs in a browser, so cost isn't a barrier. The progression runs from low-code to Microsoft MakeCode and micro:bit, which has an emulator, then Twine and finally 3D worlds in A-Frame. Looper's reasoning is that "there is ability everywhere, but access is quite limited". Teachers get a password-protected site with answers and a Common Cartridge package for learning management systems such as Canvas. Hanselman says he works with the MakeCode team.
• Local testing showed the textbook format is the weak point. Looper took the idea to the computer science teacher at her local school in Wellesley, Massachusetts, where she says computer science is often staffed from the math department, and then to a focus group of sixth graders. The students asked for video and quests, which she has begun adding on the companion website.
• Looper hasn't settled how AI belongs in teaching. She has paused CS 601, her front-end course at Boston University's Metropolitan College, while she works it out. She favours writing HTML by hand with Copilot off, and for advanced students, code-reviewing a site an agent scaffolds. In her own work she found Copilot fast for a Vue 2 to Vue 3 upgrade and Bolt.new clean, but found a fully agentic tool "terrifying" and its code sloppy. She wants Copilot as a junior engineer for toil, with the creative work kept human.

2025-04-17 · open the original ↗

• Every Wave unit backs its 60 GHz link with 5 GHz. 60 GHz is sensitive to weather and alignment, so the 5 GHz radio takes over if it drops; the host expects about 1 Gbps or more on 60 GHz but 200–400 Mbps on the backup.
• Point-to-point range and price climb together across four models. Line-of-sight ranges: Wave Pico $179, about 1.3 km; Wave Nano $279, 5 km; Wave LR $379, 8 km; Wave Pro $599, about 15 km. A link needs a pair.
• The two point-to-multipoint APs trade beam width for reach. The Wave AP Micro and Wave AP are both $499; the Micro's 90° beam covers wider, while the Wave AP's 30° beam reaches stations much further away.

2025-04-17 · open the original ↗

• Zerve pitches itself as a full-stack environment for the data science life cycle. Michaelson says that apart from hardware and data warehousing, which it connects to, Zerve covers data connection, exploration, visualisation, reports, GPU model training, self-built or SageMaker deployment and scheduled jobs. SQL, R and Python users work side by side, syncing to GitHub or Bitbucket so engineers can stay in VS Code.
• LLMs have made building so cheap that bespoke SaaS faces an existential crisis. Michaelson argues the old build-versus-buy trade-off has shifted sharply toward building, threatening SaaS vendors that sell custom use-case solutions "and then charging a mint for them", with knock-on effects for venture capital. He doesn't predict what the space will look like in a few years.

2025-04-16 · open the original ↗

• ShaderGlass overlays retro shaders on any window, invisible to capture. Hanselman demos the app, written by a self-described retrocoder whose name the captions garble, applying effects such as a Game Boy, NTSC VCR or CRT-geometry look over VS Code. Neither the podcast screen share nor a screenshot could see it, so Mark watched from his iPhone. Mark explains that it writes directly to the frame buffer, so Windows is unaware of it. Hanselman calls it "100% useless" and delightful: people should build things "for the heck of it".
• Windows Terminal compiles HLSL pixel shaders dynamically. Hanselman points the experimental pixel shader path in Terminal's default profile settings at readable HLSL files from GitHub, editing parameters such as tint colour, bloom and scan lines. He says he uses a CRT shader when talking to his PDP-11.
• Vibe-coding a fractal shader failed first, then worked. Hanselman gave Claude 3.7 thinking the Dawn shader as a base. The result failed to compile ("incorrect number of arguments to numeric type") and the session broke, and he concedes he is only at the hello-world level of writing shaders. A restart produced a working dynamic-zoom fractal shader, which he published by dragging the file into a GitHub repo and having Claude write the README, which they call shovelware.
• Copilot's code citations surface licensing questions AI doesn't weigh. The output was flagged as matching code under four license types. Mark notes that many hits were boilerplate every Terminal shader contains, and that copied "magic numbers" trace back to whoever first conceived them. One source was a 9-year-old GLSL file with no license; a function traced to a 2013 Stack Overflow answer under a "do what you want" licence, which Hanselman notes is not OSI-approved but is compatible.
• Hanselman claims AI-generated code can't be copyrighted. He says the US Copyright Office ruled that AI-generated content cannot be copyrighted, so the shader is free by definition. Only lines they modify themselves would be copyrighted, which he calls "interesting territory" for AI-assisted coding. Mark first asks whether the ruling covered only generated art.

2025-04-16 · open the original ↗

• Serverless blocks make parallel execution free of extra code or cost. Michaelson says Zerve's new "fleet" feature, launching the week of recording, fans code out across serverless compute, so a thousand slow LLM calls can run at once instead of in series. Because each block in Zerve's DAG already spins up its own serverless compute, parallelising needs no pool management and, he says, costs no more since total compute is the same.
• A shared DAG canvas does for notebooks what Google Docs did for documents. Code lives in the graph's nodes and data frames flow along its edges, with forks and merges, Michaelson explains, so several people can run Python, R and SQL blocks in one canvas simultaneously. He contrasts Jupyter, where metadata changes cause merge conflicts and "final final" file names; the host supplies the Google Docs analogy.
• Low-code tools hit a ceiling, so value comes from people who code. Michaelson, who spent about seven years at DataRobot, says the emperor of 2010s low-code/no-code "has no clothes", since complex problems exceed such tools, which is why DataRobot and Databricks added notebooks. Zerve, he says, remains code-first, though LLMs have turned coding into describing what you want and then tweaking the result. The host relays Zerve's claim of cutting cycle times by up to nine times; the clip has no evidence for the figure.

2025-04-15 · open the original ↗

• Cheap in-house building with LLMs threatens many SaaS vendors. Michaelson argues the build-versus-buy calculation has shifted so far that firms selling custom use-case software face an existential crisis, though nobody will rebuild something like Salesforce, and he separates "dime a dozen" generative AI bolt-ons from deeply integrated agents. The host sets this against Klarna's and Satya Nadella's similar views and Zoho's CEO's counter-argument that AI will spawn new vertical SaaS, and notes a friend who cut a $20-a-month ChatGPT Plus cost to $2 by calling the API through his own interface.
• Zerve's serverless blocks make parallelism and collaboration nearly free. Michaelson explains that code lives in the nodes of a DAG canvas, with data frames passing along its edges, and each block spins up its own serverless compute. That lets several people run Python, R and SQL in one canvas at once, and lets the new "fleet" feature fan out, say, a thousand slow LLM calls with no extra code and, he says, no extra cost.
• Zerve lets companies choose where prompts go to protect their IP. Michaelson says OpenAI's models are the best in his experience, though he distrusts LLM benchmarks, and sending prompts there is unavoidable with them. Zerve also offers AWS Bedrock and self-hosted Hugging Face models, which can be fine-tuned, and Zerve itself is self-hosted. Beyond prompt-driven GenAI blocks, its AI assistant is an agent that plans, creates and connects blocks from a natural-language request.
• Data scientists should ship software, not prototypes for engineers to recode. Michaelson calls the handoff to developers a risky dependency, since they may misunderstand the model. Zerve gives each canvas a Docker container and serialises every block's variables, so an API route can point straight at a trained random forest and deploy on Lambda-style serverless compute, with a handoff path where data scientists aren't allowed to deploy.
• LLM committees are his answer to bias and error in AI judgment. In Michaelson's test, four LLMs answered and then ranked responses to about a hundred SAT essay questions; rankings were "wildly unstable" across reruns, but no model favoured its own answers. He suggests cheap models screening answers and escalating doubtful ones to expensive models, and warns that a model that gets a problem badly wrong first time is rarely correctable.

2025-04-14 · open the original ↗

• Dell estimates its toolkit cuts NPU app delivery from six months to six weeks. Gupta says that by Dell's estimate a typical, not very complex AI PC app such as a RAG chatbot takes about six months from discovery to fleet deployment, covering model and device selection, runtime compatibility, on-device model operations and manual deployment. Dell Pro AI Studio, in early access at recording and due fully mid-2025, would bring that under six weeks, about a 75% reduction, leaving developers to spend time on picking the use case and validating output accuracy.

2025-04-13 · open the original ↗

• Today's NPUs comfortably run quantised models of about 8 billion parameters. Gupta puts acceptable on-device output at around 15 to 20 tokens per second, which he says rules out going much beyond an 8 or 9 billion parameter model with quantisation. NPUs passing 40 TOPS and smaller LLMs growing more capable are, in his view, the inflection that makes local inference practical.
• Dell-validated models end trial-and-error pairing of models and machines. Gupta describes developers otherwise iterating through 14 and 30 billion parameter models and memory upgrades only to land at two tokens per second, a deterrent to building for the NPU. Dell Pro AI Studio, a developer and IT toolkit rather than a preinstalled app, offers open, permissively licensed language, speech and vision models tested against Dell systems across silicon vendors, published on the Dell Enterprise Hub on Hugging Face.
• Enterprise readiness and on-device middleware are what local-chat demos lack. Gupta says enterprises need attestable models, control of data in and out, and lifecycle management of permissions, updates, deprecations and telemetry, which a Dell management portal provides. On-device middleware detects the silicon, pulls the intended model from a registry, loads, unloads and verifies it, replacing hundreds of lines of code developers would otherwise write and rebuild with each model change.

2025-04-12 · open the original ↗

• Build anywhere, but integrate and test AI PC apps on Windows. Gupta argues that because NPU inference apps will mostly be consumed by Windows users, developers can do much of their work in an IDE of choice on Unix-based machines, but must integrate the model into the app, set up the local host and verify it runs on a Windows device, since that is where the app will run.

2025-04-11 · open the original ↗

• AGI needs planning over an unlimited horizon, which LLMs lack. Burkov, in an episode 867 clip, contrasts crows and monkeys that plan a day or two ahead with humans who plan for generations, and says LLMs are reactive: an "agent" has only the agency its system prompt gives it. He doesn't hazard a timeline, likening AGI to teleportation. The host suggests it may need brain-like modules such as a prefrontal cortex rather than a scaled-up transformer.
• Models can't tell what they don't know, and fine-tuning hacks don't fix it. Burkov says o3-mini's visible reasoning admitted it knew nothing of his new book, released after its cutoff, yet its answer praised the book as if fact. He says Qwen models fine-tuned to say "I don't know this person" denied knowing him, then named him as author of his first book.
• Doubling down on embodied experience is the human edge over AI. In an episode 873 clip, a digital-twin entrepreneur named Natalie argues AI is part of our cognitive evolution as language is, but not all of us. Embodied, serendipitous experience, like the host's childhood browsing of dictionaries, builds insight AI can't replicate, so people should collaborate with AI and use it to free themselves to be more human.
• The AI stack looks different to executives and developers. Richmond, a MongoDB software engineer, says in an episode 871 clip that executives see three layers: application (such as Cursor), tooling (such as MongoDB) and compute (such as Nvidia). Developers need more, starting with programming language, since many libraries exist only in Python or JavaScript, then model provider and database.
• Google's tuning playbook tries to make neural-network training systematic. Varun, a former Google engineer and first author, says in an episode 869 clip that the playbook moves training from "alchemy" toward chemistry. Architecture-agnostic, it helps practitioners choose the next experiment given a compute budget. It was released as a Markdown file on GitHub so others could fork it, and had 28,000 stars at recording, per the host.

2025-04-10 · open the original ↗

• Missing data, not models, is the Global South's AI barrier. Kallott, founder and CEO of Amini in Nairobi, says 2.6 billion people are unconnected and so "invisible" to training data, and Africa has more than 2,500 languages, many only spoken. The continent has only 1/8 of the minimum density of meteorological stations. Her team set out to build an AI company for Africa, found there was no data, and pivoted to data infrastructure, starting with environmental data.
• Global earth-science models lose accuracy on African ground. Kallott contrasts large US monocrop farms with African farms of around one square kilometre growing several crops, so open satellite imagery at 10 metres has to be super-resolved toward one metre. Readings are calibrated against ground truth, such as soil samples and yields collected by data partners. Amini worked bottom-up, starting with one region of Kenya, because she says East Africa has little cloud cover and the West a lot.
• Amini built its own geoprocessing engine to beat cloud cover. Over about a year and a half it built a Google Earth Engine–style system that layers multispectral, hyperspectral and synthetic aperture radar satellites and applies cloud masking. It also takes in drone imagery, tabular data and IoT sensors, including weather stations built into a startup's e-bike chargers.
• Revenue comes from insights; models and community access stay free. Kallott says customers are global exporters of agricultural and mineral commodities wanting first-mile data, and governments mapping their resources, while developers never pay. Version 0 of the foundation model went on Hugging Face the month before, with the next version in training and a first ICLR poster accepted. Amini runs hackathons on Zindi, such as telling cocoa from coconut trees in Côte d'Ivoire. Expansion to the Caribbean and Southeast Asia rests on her finding that about 70% of work by agro-ecological zone carried over.
• Africa must become "AI-ready" before it can leapfrog. Kallott claims Amini is the only company doing this and wants to build "an ecosystem, not a product". She compares it to Palantir digitising government data more than 10 years ago, though not in defence. She argues the work is best done by teams "born from within", who build with a UPS and generator through power cuts and train foundation models on office workstations.

2025-04-10 · open the original ↗

• Local accelerators give low latency for persistent real-time work. Gupta's own mnemonic starts with "accelerated": an on-device NPU suits latency-sensitive, always-on tasks such as translation, transcription and captioning.
• On-device AI can learn an individual user's style. For "individualized", Gupta cites a healthcare customer auto-generating ER physicians' reports from their diagnoses; a physician reported that as the model saw his edits it began to sound like him, freeing time for patients.
• Data never has to leave the device. For "private", Gupta argues local models avoid sending confidential or PII-laden data even to a private cloud tenant.
• Offloading inference to PCs cuts data-centre and per-token costs. For "cost", Gupta says a financial-services customer found 15% of its data-centre compute going to developers' code generation and completion, work he believes NPUs can take over without H100s, and cloud inference bills per token. He concedes cloud inference costs are falling, but argues for "the right engine for the right workload". Edge vision cases include real-time factory defect detection and an insurer classifying damage photos into rough cost sizes to get adjusters "maybe 70% of the way there".
• Dell works with chip partners so models run across NPU silicon. Asked about chips from Intel, Qualcomm and AMD, Gupta says the NPU architecture is so new that Dell is preparing models with its silicon partners. This lets developers target different chips without starting over, one reason Dell built Dell Pro AI Studio.

2025-04-09 · open the original ↗

• NPUs are power-efficient matrix-math engines new to ordinary PCs. Gupta says NPUs reached the market about a year before recording, first in Intel's Meteor Lake chipset, joining the CPUs and integrated or discrete GPUs PCs already had. Being purpose-built, almost hard-coded, for the matrix multiplication and addition underlying AI, they use far less power than running models on a CPU or GPU, which would drain a laptop battery.
• For now NPUs are for inference, while GPUs still build the models. Gupta expects data scientists and developers to keep using GPU workstations for training and fine-tuning. NPUs will serve the knowledge workers who consume the results, from assistants answering questions over a knowledge base to AI embedded in workflows and, he foresees, multi-agent systems, without tanking battery life.

2025-04-08 · open the original ↗

• Clearly IP's free Open Page module unlocks formerly paid paging features. It works with FreePBX and Tango PBX version 17; install by switching to the Clearly IP repository, then running fwconsole ma downloadinstall openpage over SSH, which adds an advanced paging section to page groups.
• Pages can carry a lead-in tone and go out by multicast. A prepend announcement plays a system recording, such as a chime, before the live page; one or more multicast addresses let large numbers of phones be paged without consuming phone-system resources.
• Valet and scheduled paging suit noisy sites and schools. Valet paging records the whole message first and can queue several pages to play in turn; scheduled paging runs pages on a recurring timetable like a school bell schedule, with three bell sounds included or your own uploaded.

2025-04-08 · open the original ↗

• Backlogs cost work without delivering customer value. Henriksen, who runs the technical side of a legal-tech startup, made the case in a Copenhagen DevFest talk pairing "ditch your backlog" with Shape Up, which his startup uses. He says it drew workshop requests from companies. Raw user stories are "placeholders for communication" whose context has faded or gone by the time anyone discusses them, and grooming and reprioritizing them is effort taken from the customer.
• Backlogs also trap teams in cognitive biases. He names recency bias (newer items get more attention), the IKEA effect (people overvalue items they put work into) and the sunk-cost fallacy (we specced it, so we must build it). His answer is that nothing is more wasteful than building the wrong thing, which also costs the chance to build the right one.
• Every choice should have a stated reason, including keeping the status quo. Henriksen, who has taught software engineering at a local university for about 12 years, points to the repository pattern (which he dislikes) and Scrum as things adopted on a book's authority. Appeal to authority is fine for a beginner, he says, but not with 5–10 years of experience. His team started with MediatR, then weighed Wolverine and the "Critter Stack" and Steve Smith's endpoints library, and kept MediatR for reasons it stated. Keeping something because there is no time to change is also deliberate, he says, as long as you see that you are making a decision.
• He cites Microsoft's Go-based TypeScript compiler as a deliberate choice. He reports that the rewrite promises up to 10 times the speed. By his account, Anders Hejlsberg explained that the team evaluated many languages and chose Go for native code and a style close enough to TypeScript's for a near one-to-one port, where C# would have meant a new paradigm. Henriksen adds that the developers were TypeScript people, so they had to learn something either way.
• Unexamined stacks lead into the "COBOL trap". He gives COBOL banks whose developers are retiring, AngularJS, .NET Framework and a system built on Delphi as examples of postponing a problem that hits hard when it comes. Staying put can be fine if the reasons are named. Each reason is smaller than the whole problem and can be removed one at a time toward a destination. He applies the same deliberateness to jobs, promotions and friendships.

2025-04-08 · open the original ↗

• NPUs are inference engines for end users, not training hardware. Gupta, of Dell, says NPUs arrived about a year before recording with Intel's Meteor Lake and are near-hard-coded for matrix math, so they run AI far more power-efficiently than a PC's CPU or GPU. GPUs remain for training and fine-tuning. He puts acceptable output at 15 to 20 tokens per second, which limits today's NPUs to quantised models of about 8 or 9 billion parameters, now capable enough to make local inference practical.
• His "AIPC" mnemonic makes the case for on-device AI. Gupta's letters stand for accelerated (low latency for translation or captioning), individualized (an ER report generator that learned a physician's style), private (data never leaves the device) and cost. On cost, a financial-services customer found 15% of its data-centre compute going to developers' code completion, which he thinks NPUs could absorb. He concedes cloud inference is getting cheaper and that not every use case belongs on the device.
• Dell says its toolkit cuts NPU app delivery from six months to six weeks. Gupta describes Dell Pro AI Studio, in early access at recording and due mid-2025, as three parts. Validated models on the Dell Enterprise Hub on Hugging Face end trial-and-error pairing of models and machines; a management portal adds attestation, permissions, updates and telemetry; on-device middleware detects the silicon and loads the right model. By Dell's estimate, a typical app such as a RAG chatbot goes from discovery to fleet deployment about 75% faster.
• Build on workstations, scale on servers, deploy to AI PCs. Gupta advises embedding data scientists with the people whose problem needs solving, then moving production inference to AI PCs if H100-class hardware isn't needed. His example is a healthcare customer that trained a custom vision transformer on its own radiology images and reports physician-level diagnostic accuracy. It now runs in a HIPAA-compliant data centre, and the customer wants to bring it to AI PCs for doctors worldwide.
• On-device agents await small reasoning models; hybrid compute comes sooner. Gupta forecasts agents acting on users' behalf within set parameters, such as buying concert tickets on approval, but says reasoning models must first shrink enough to run locally, which "might take some time". He also expects workloads to shift seamlessly between device and cloud as compute needs demand.

2025-04-07 · open the original ↗

• Merck's display dominance plateaued, so it moved into semiconductors. Beckmann, who heads Merck KGaA's 8,000-employee electronics business, says its late-1960s and 1970s liquid-crystal research took decades to commercialise. At peak, roughly 2005 to 2010, it held 80% of the liquid-crystal materials market. Recognising that every innovation ramps then plateaus, it entered semiconductors around 2014 and built a leading portfolio through acquisitions including AZ, Sigma-Aldrich's high-tech business, Versum and Intermolecular.
• Incumbents' worst innovation mistake is assuming yesterday's success lasts. Beckmann argues that innovating inside an established company is far harder than in a startup because "you are in your own way", and calls believing that what worked yesterday will work tomorrow probably the most fatal mistake in business. He traces Merck from its 1668 founding after the Thirty Years' War to what he says was industry's first quality promise to customers around 1850.

2025-04-06 · open the original ↗

• High-bandwidth memory depends on hundreds of specialised materials. Beckmann says HBM needs the best possible DRAM performance, and DRAM capacitors rely on high-k materials where he says Merck leads. Spin-on dielectrics are among many promising materials, but with hundreds or thousands of precursors drawn from 80% of the periodic table, he declines to single out one as most important.
• Light-based data transfer promises speed and energy gains. Beckmann says photons move data faster than electrons and without the same heat, and that Merck reorganised its electronics sector the year before recording to create an optronics unit built on its display-era expertise in generating and manipulating light.

2025-04-05 · open the original ↗

• Heterogeneous integration scales performance beyond shrinking transistors. Beckmann describes it as the step between front-end transistor making and back-end packaging: combining several dies into one system, stacking dies into memory stacks or placing memory beside a GPU, as in CoWoS-style structures, to shorten data paths. It requires far smaller structures, better heat removal and power optimisation, and front-end-like precision, which opens room for materials and metrology innovation.
• Merck uses AI to skip unnecessary materials experiments. Beckmann says its "materials intelligence" approach applies AI across millions of candidate options to predict how a material will behave in a customer's setup, rather than improving properties sequentially by experiment.

2025-04-04 · open the original ↗

• Ubiquiti's USB-C 10 GbE adapter delivered near line rate in testing. Plugged into a Mac Mini's USB4/Thunderbolt port and cabled to a UCG-Fiber, it measured 9,750 down and 9,776 up (the captions give no unit) against an OpenSpeedTest server on a Minisforum MS-01; MSRP is $199.

2025-04-04 · open the original ↗

• smolagents keeps agent logic small enough to read and tweak. The host describes Hugging Face's new Python library as giving any LLM the ability to use tools, call APIs and reason over multiple steps. Its core logic fits in a few thousand lines with minimal abstraction over raw Python, and it succeeds the transformers.agents module, which it will eventually replace.
• Its agents act by writing Python code, run in a sandbox. Rather than expressing actions in JSON or text, smolagents' code agents generate Python that is then executed, which the host says gives them Python's full flexibility, with sandboxed execution as the safeguard. A conventional tool-calling agent is also offered.
• Tools are plain decorated functions, shareable on the Hub, with any model. Any Python function becomes a tool with an @tool decorator, type hints and a docstring, and tools or whole agents can be shared and loaded from the Hugging Face Hub, "like an app store". Models can be open ones via Transformers or the inference API, proprietary ones from OpenAI or Anthropic via LiteLLM, or local ones.
• It lowers the bar for agents in business, data science and research. The host lists research question-answering, as in Hugging Face's demo of a leopard running across a Paris landmark, one-line CLI trip planning, browser automation via its web agent, and on-the-fly pandas analysis. He argues firms can prototype agents in a few lines and keep data in-house with local models, and researchers can benchmark small open models against large proprietary ones. He states the episode is not sponsored by Hugging Face.
• Heavier frameworks beat it for scale, data and multi-agent work. By the host's comparison, LangChain surpasses smolagents in scalability, memory management and integrations, and LlamaIndex in large-data retrieval. Pydantic AI offers type-safe validated outputs, while Microsoft AutoGen and CrewAI are built for coordinated multi-agent systems, where smolagents focuses mainly on single agents.

2025-04-03 · open the original ↗

• Swap linear goals for experiments wherever life allows. Le Cunff, a neuroscientist, founder of Ness Labs and author of Tiny Experiments, says linear goals suit narrow, predefined outcomes such as saving for a mortgage. Elsewhere she would start from a question or hypothesis you are curious enough to test, rather than a fixed outcome you tie your happiness to.
• The obsession with finding one purpose backfires. In her research for the book she found that mentions of "finding your purpose" in books rose 700% over two decades. She argues it puts all of a person's self-worth in one basket, citing founders who suffer chronic anxiety and depression after a startup fails, and that people who have "found their thing" describe a messy process, not a plan. Hanselman blames social media, and she calls it "a giant artificial leaderboard". She doesn't tell people to stop scrolling, since she still does it herself. Instead she suggests replacing some of it with sharing something useful once a week.
• A tiny experiment is an action plus a duration, reviewed honestly. Her own example: during a book launch with recordings from 8 a.m. to 8 p.m., she took a 20-minute walk daily for 20 days, then judged by felt signals such as mood and sleep whether to continue, tweak or stop. A weekly YouTube video for about three months drew good subscribers and comments, but she dreaded making them and quit. She says this kind of trial should come before committing to a habit, and that the popular 66-day habit figure oversimplifies the research.
• Learning in public is a spectrum, starting with one person. Le Cunff says it can be one friend or a temporary WhatsApp group exploring a new tool for three weeks, not a large following. It is optional, but she says a support system, like a marathon runner's team, is how tiny experiments grow into ambitious ones.
• Procrastination is a signal to diagnose, not beat with willpower. Her "triple check" asks whether the block is in the head (should this be done at all?), the heart (is it enjoyable?) or the hand (are the tools and skills there?). Applied to Hanselman's TED talk outline, stuck at 60% for three months, she diagnoses heart and hand. Her remedies are intentional imperfection, working alongside a fun friend, and asking someone who has given a TED talk for the two things that matter.

2025-04-03 · open the original ↗

• A chip takes about 1,400 repeated steps from blank wafer. Beckmann explains that leading chipmakers build transistors and power and data lines on 300 mm silicon wafers, the size of a vinyl record, cycling through lithography, patterning, deposition, cleaning and planarisation until the wafer is cut into centimetre-scale dies.
• Chipmaking draws on about 80% of the usable periodic table. Beyond photoresists for lithography, Beckmann says the specialty materials for etching, deposition and planarisation span about 80% of the non-radioactive periodic table. His examples mostly concern logic chips such as CPUs and GPUs, which he calls the main driver of compute, as distinct from memory.
• His route to leading Merck's electronics arm ran through IT and HR. Beckmann studied computer science with a semiconductor focus and worked on chip design in the late 1980s, then spent years in Merck's corporate IT, ran a country sales organisation, and served as board member for HR during a period of transformation before being asked to head electronic materials.

2025-04-02 · open the original ↗

• Internal audiences allow a candour external talks can't. At Microsoft's internal TechReady conference, Hanselman gave a talk arguing "we're selling Copilot wrong" and found the openness so freeing that he is now saying more of it publicly. Mark describes giving an Azure strategy talk on the hyperscaler landscape there that he could never give externally because it is business strategy and opinion.
• Mark publicly called for treating C and C++ as deprecated. In a 2022 tweet he argued Rust was ready and the industry should shift to memory safety, after reviewing trends showing 70% of vulnerabilities were memory-safety problems. He says he posted it as a technical opinion without press clearance and got a Teams call the next morning asking whether he really believed it. The stance built the reputation that led to his recent keynote at Rust Nation UK.
• Microsoft tolerates employee posting, but discretion still applies. Hanselman says Microsoft is far more relaxed than Google or Apple, where friends were told to stop blogging or YouTube. Mark says he has told team members that social media is the wrong forum for some points. Hanselman's team rule, "don't dunk on other teams", comes from a tweet about Windows 10 printers that brought three VPs down on him; the bug got fixed, but he says he should have asked internally.
• Being inside lets them get Windows bugs fixed directly. Mark recounts a printer failure that took a driver reinstall to clear and says the Windows printing team took it up; they point to the universal printer driver as the fix for new printers. Hanselman says a regression in how USB device INF files were set broke Zune on Windows 11, which he found while preparing Zunes for Marvel's Guardians of the Galaxy 3, and that the fix has rolled out.

2025-04-02 · open the original ↗

• Neuromorphic chips may pay off first in inference, within about five years. Beckmann argues neuromorphic architectures could matter even more for cost-efficient, scalable inference on everyday devices than for training, given that data transfer is a key optimisation target. He calls their use in general compute "a five years plus question rather than a 10 years question", but says it probably won't reach human-brain efficiency; the light-bulb-powered brain comparison is the host's.
• Data-centre AI already scales on several fronts at once. Beckmann says chips such as Nvidia's Hopper and Blackwell combine transistor shrinkage via EUV lithography and gate-all-around nanosheet designs with CoWoS-style integration that places memory beside compute. The same latest technologies are reaching smartphones, laptops, cars and production systems.

2025-04-01 · open the original ↗

• Chip performance now scales along three dimensions, not just shrinkage. Beckmann, CEO of Merck KGaA's electronics business, distinguishes "more Moore" (smaller transistors via EUV and gate-all-around nanosheets), "more than Moore" (heterogeneous integration such as CoWoS, placing memory stacks beside GPUs as in Hopper and Blackwell) and post-von Neumann architectures. He expects neuromorphic computing in general compute in "five years plus" rather than 10, though not at human-brain efficiency. Superconducting quantum approaches need near-zero-kelvin cooling and remain a tiny market; Merck's Intermolecular facility in San Jose is exploring qubit materials with partners.
• Co-optimising materials, delivery and inspection is meant to compress years-long cycles. Beckmann says chipmakers' yield problems are solved fastest by integrating chemistry, physics, electrical testing and chemical delivery systems. That is the logic for acquiring UnitySC, whose high-resolution visual inspection checks 2.5D and 3D structures such as through-silicon vias. Merck calls its use of AI, including on metrology data, to cut unnecessary materials experiments "AI for AI".
• AI drives leading-edge chips while the rest of the industry stays in a cyclical slump. Beckmann says volume still depends on smartphone and PC replacement, which remains low as people keep COVID-era devices into a fourth or fifth year, and AI data-centre demand "doesn't bail everything else out". Because fabs run either at full capacity or idle, the industry is inherently cyclical; after semiconductor-materials growth above 20%, he says Merck stays humble, knowing shrinking years follow.
• Good AI regulation protects data contributors without slowing permits. Beckmann argues frameworks should ensure those who make data available get returns when it feeds LLMs, since that stability helps investment. But he says that needing about a year to get a permit to make a new material in Darmstadt hurts innovation, and that Europe is not leading on approval speed.
• Every innovation plateaus, so incumbents must switch gears deliberately. Beckmann recounts that Merck's liquid-crystal research from the late 1960s led to an 80% share of liquid-crystal materials at the flat-panel peak around 2005 to 2010. It then moved into semiconductors from 2014 through acquisitions. He calls assuming yesterday's success will last "probably the most fatal mistake you can do in business".

2025-03-31 · open the original ↗

• Experts can rent out vetted AI twins of themselves. Monbiot says she now advises a startup (spelled on air W-I-Z-L-Y) that admits experts selectively, lets them train a twin on knowledge they control, use it to generate leads, and monetize it in consulting calls once they trust it to represent them.
• A twin's value is augmented skill, not just bought time. Because the twin is cheaper, always available and speaks any language, Monbiot argues it can serve clients abroad at a price point they can afford, which she calls the "10x" of the idea.
• A CEO's twin can carry company culture across an organization. Monbiot describes building one for a healthcare CEO, first inside Slack, where employees can consult it on his values, have it interject, and get feedback on work, including staff at a company recently acquired; voice and other modalities are to follow. The host flags liability as an open question when people act on the twin's advice.

2025-03-30 · open the original ↗

• Each step from books to Google to LLMs removes cognitive work. The host argues that looking things up in a dictionary or encyclopedia exposed him to incidental information that built a web of connections and aided memory, while a search page surfaces ads and an LLM output does the comparing for you.
• Humans should double down on embodied experience to stay competitive with AI. Monbiot holds that AI is part of human cognitive evolution the way language is, but not all of us: being embodied and embedded in the world, noticing things and making real-world connections is the human advantage AI lacks. Her prescription is to use AI to solve big problems and free ourselves "to be more human".

2025-03-29 · open the original ↗

• Outsourcing whole cognitive processes to LLMs makes us dumber, unlike Google. Monbiot answers commenters who compare the worry to early fears about Google: search returns facts that feed into your own reasoning, whereas an LLM can tempt you to hand over the entire process without reviewing or challenging the result. She expects this to change how children are educated and how adults reskill to avoid that trajectory.

2025-03-28 · open the original ↗

• Technology adoption in medicine "flips" expectations. Lee's example is ultrasound: proposed as a diagnostic tool in the late 1950s, feared, then standard of care about a decade later, to the point that refusing it now would look like malpractice. He concedes skills are lost along the way, but says the depth of debate among leading thinkers, which he compares to the one after the human genome was mapped, makes him optimistic.
• Reaching a capability first costs roughly ten times more per tier. Lee puts a GPT-2-level model at about $5–10 million of compute, GPT-3 level at $50–100 million, and GPT-4 level near $1 billion, in any domain. Once someone shows it can be done, followers get there "smarter and a hell of a lot cheaper", and he asks whether the premium on being first still holds.
• AI will soon build things that work but no human understands. Lee predicts that in "much less than 5 years" an AI will produce a mathematical proof in a proof language such as Lean that type-checks as valid but defies even the smartest humans' comprehension. He says that would be his first mark of AGI or superintelligence.
• Patients rated GPT-4's replies far more empathetic than doctors'. Lee cites a UC San Diego and Stanford paper, published a month after his book, in which GPT-4's answers to patient emails were equally accurate and judged more empathetic by a factor of 9 to 1 in a blind test. His reading is that a frazzled doctor writes two or three sentences while the AI writes paragraphs and recalls personal details, which he calls both interesting and disturbing.
• AI for science is stalled at GPT-2 level for lack of data. Lee says Microsoft Research's models for weather, proteins, molecules and materials are published in journals and released on Azure Foundry simultaneously, but there is "no internet of molecular dynamics simulations". He predicts the first practical scalable quantum machines in 2025, and wants them first used to generate large sets of perfectly labelled training data.

2025-03-28 · open the original ↗

• Image-based, A/B-updated Linux desktops give Chromebook stability without the restrictions. Garrison runs Bluefin, built on Fedora Silverblue, where rpm-ostree works like "git for your file system": a full distribution upgrade is just a reboot, and he can bookmark an OS version and fall back to it. He installs command-line tools with Brew and GUI apps as Flatpaks, and sees the read-only system as a security practice moving from servers onto the desktop.
• Linux already runs the media industries' workstations. Garrison managed Linux at Disney Animation, where all the artist workstations ran Red Hat Linux and the company ran on Google's web apps rather than Microsoft Office, and he says DreamWorks and ILM run Linux too. Having also managed thousands of Windows machines at a university, he says Windows "gets broken just the same".
• Wayland's strictness hurt, but its sandboxing is deliberate. Garrison says the move from X11 was rough because Wayland refused broad compatibility and much software was never updated. In his view the isolation is there for a reason, though it still costs him, for instance when he cannot drag a file from an NFS share into a sandboxed Slack.
• Owning more of your stack is worth the friction. Garrison has switched to Firefox and is "de-Googling", and finds many web apps are Chrome-specific rather than web-standard. He rips DVDs into Plex and runs local home automation that keeps working when the internet is down. Hanselman calls pushing back against a browser monoculture noble rather than self-inflicted.
• Give non-technical users a Chromebook, or an appliance-style distro. Garrison's first answer for family is a Chromebook. For Linux he now recommends Bazzite or Bluefin, because updates arrive automatically on reboot, and he says Linux Mint, which he once promoted, lacks a good update story.