Souk's Notes - Unofficial Summaries
Polis Pandit$What Warren Buffett Missed on Tariffs
- This post includes paid content that is not available publicly. See the original source.
Crosstalk SolutionsCrosstalk's Travel Tech Setup for 2025 #tech
- The GL.iNet Slate 7 was "absolutely perfect" as a travel router. On its first trip, at RSAC 2025, the host ran it as a wireless repeater because his hotel room had no Ethernet port, with his laptop hardwired to it through an Anker multiport hub; the kit also includes a YubiKey and a 1080p portable second monitor.
- The host's roughly four-year-old Samsung laptop needs replacing. DaVinci Resolve crashed several times on the trip while editing in only 1080p, and he is looking for a laptop for video editing.
Stock Tok$Investing For Beginners - 10 Short Videos
- This post includes paid content that is not available publicly. See the original source.
Dr. Jon KrohnThe Future of AI: What’s Just Around the Corner (with Logan Lawler and Sama Bali)
- 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.
Dr. Jon KrohnInside NVIDIA’s AI Stack: CUDA, TensorRT, and NIM Microservices
- 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.
Mr. GlobalNatural Gas Will SKYROCKET. As Will Your Utilities.
- Falling oil prices cut US natural gas supply. He says a little over 25% of US natural gas output is associated gas from oil wells, so it falls with oil production; historically that mattered little, but he argues it now does because US gas inventories have been tightening and will keep tightening.
- LNG export growth is what keeps domestic inventories tight. He puts US exports at over 11 billion cubic feet a day, with export capacity at record levels and on track to more than double by 2028 and triple sometime after. Because gas is priced locally, he says even today's low export prices are about twice domestic ones, so companies will export all they can.
- Lower crude will shrink US oil output sharply this year. He expects marginal, barely economical old wells, of which he says there are thousands, to shut first, then rig counts to fall below what decline rates require. His forecast: if oil reaches $50 a barrel, US production falls by about 1 million barrels a day by year-end, driving gas prices up drastically.
- Higher utility bills will outweigh any savings on gasoline. He cites natural gas demand growing about 4% this year, which he calls huge, alongside falling supply and doubling exports. He predicts a gas crunch and much higher utility bills, saying Americans are already downsizing homes over bills and that saving 50 cents a gallon on gas is no reason for excitement.
Dr. Jon KrohnHow to Get an NVIDIA Blackwell Chip in Your Desktop PC for AI Development
- 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.
Crosstalk SolutionsUniFi Pro Max Rack Cost Breakdown! #networking #rsac2025 #tech
- 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.
GCastCreating a Chatbot with the Question Answering feature of Azure Language Service [GCast 194]
- 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.
Stock Tok$Stock Watchlist - 02 May 2025
- This post includes paid content that is not available publicly. See the original source.
Dr. Jon Krohn884: Model Context Protocol (MCP) and Why Everyone’s Talking About It (@JonKrohnLearns)
- 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.
Scott HanselmanMaking SNES ROMS with C# and Matthew Shapiro's DotnetSnes
synopsis
Matthew Shapiro, a staff video architect at GameChanger, has a hobby project that lets C# programs become Super Nintendo ROMs. A Reddit post put it on the Hacker News front page for a day and brought its first outside contributor. The pipeline translates .NET IL to C, then uses an existing SNES C toolchain. He sees it as a general route to running C# anywhere a C compiler exists, and he plans to target the PlayStation 1 next.
Origins. The project began on parental leave, optimizing a friend's retro game engine on Wilderness Labs Meadow hardware, which led to writing a small GPU in C for an ESP32. Wanting C# there instead, and noting that Nim compiles to C, he tried translating MSIL to C. Its first target was eBPF, kernel programs that cannot allocate memory, which matched what his transpiler could handle. A friend then suggested the SNES.
How it works. DNTC builds a dependency graph from a root function and converts the stack-based IL into C statements. Shapiro said working ahead of time showed him how hard the JIT team's optimizations are. There are no reference types yet because there is no garbage collector, which matters little on a console with no dynamic allocation. Attributes rename functions and bind to native C and assembly symbols. Plugins adapt the output to PVSnesLib's non-C99 C. Everything links into one binary, so native calls are free.
Shoulders. Shapiro credits Mono.Cecil, PVSnesLib and the older SDK it came from, and the Tiled map tool, which gives collision handling. Hanselman built the Mario sample in WSL in about 10 minutes. Shapiro said someone has run Breakout on real hardware.
- 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
mainand 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.
Mr. GlobalTrump Threatens China With Sanctions
- Secondary sanctions on Iranian oil escalate the China trade war. He reports the administration's announcement that any country or person buying Iranian oil or petrochemicals faces immediate secondary sanctions and a bar on US business; China buys the vast majority, and it came hours after Defense Secretary Pete Hegseth's warning to China on X. Oil rose modestly after heavy overnight losses.
- The sanctions give Washington no leverage over Beijing. He argues existing tariffs have already all but ended US-China trade, so being cut off from US business means little to China, which he expects to keep buying Iranian oil; the US has already played its hand.
- He suspects the Iran threat is an oil-price manipulation. His reading: a tanking economy has pushed oil down, which defeats "drill baby drill"; an attack on Iran would send oil north of $100 a barrel and spur drilling, but would also end any promise of lower gas prices and compound the economic damage.
- Striking Iran would likely start the largest war in our lifetimes. He sees no scenario in which Iran does not retaliate in many places across the Middle East, turning it into a broad conflict.
GeoHussarWe are closer to World War III than we’ve been in a while, but the chances are still quite lolw
- Today's feared adversaries lack the parity that made world wars possible. He argues both 20th-century world wars were fought among the most developed industrial and military powers, with both sides having a fairly equal chance of winning — a condition he implies North Korea, China, Iran and Russia do not meet against the US side.
- NATO's collective response makes attacking a member prohibitive. He cites the US's 750 military bases worldwide and by far the largest military budget, and says an attack on any NATO member would draw all 32 members in as one. The captions end before the "most pressing reason" he promises for the close.
GeoHussarAmerica’s Economic Output Shrank #econ #economy #gdp #tariffs #history #recession
- US GDP contracted last quarter, one step from an official recession. He reports a 0.3% contraction in the first three months of the year (the captions say 2024, in a video posted 1 May 2025), and says a second consecutive drop would make it a recession officially.
- There is no evidence of a recession yet. He rejects claims the US has been in recession for a couple of years, pointing to no definite GDP data and no spike in unemployment. He expects much more layoffs if the next three months also show negative output.
- Tariffs are the sole cause of the contraction. He says no other explanation holds weight, and that other countries are now erecting trade walls against the US as well. He compares the scale to the 1930 Smoot-Hawley tariffs, which he says exacerbated the Great Depression.
Stock Tok$Blow Off Top? Or Bull TRAP? 🪤
- This post includes paid content that is not available publicly. See the original source.
GCastUsing the Sentiment Analysis Feature of the Azure Language Service [GCast 193]
- 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.
Dr. Jon KrohnWhat Are Dell Pro Max PCs? (with Logan Lawler and Sama Bali)
- 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.
AI Analyst LabThe 1% Skill AI Can't Replace - How to Stay Relevant in the Age of AI | Ep25
synopsis
Luis Serrano, founder of the Serrano Academy YouTube channel and formerly at Google, Udacity, Apple and Cohere, makes one argument across this episode of the Data Neighbor podcast: people know far more than they think, and the human part of any skill is what is left once a machine has done the rest. He ends on two prescriptions. Try to get a computer to do 100% of your job and see where it falls short, because that remainder is you. And value malleability, the ability to keep learning new things, over stored knowledge.
Why people think they are bad at math. Serrano describes himself as a slow "understander" who has to tell himself stories before abstraction makes sense, and says that is why he teaches. He was failing grade-8 math in Colombia when he took a puzzle-based olympiad exam to skip class and placed first in the country. His conclusion is that school turns students into instruction followers, and that most STEM subjects are taught in the language of abstraction, which suits a small minority. The examples are concrete ones. Everyone draws a parabola for a falling tennis ball. Children splitting candy know a third beats a quarter. Music is as complex as math, yet nobody says they are bad at music, because it is introduced through songs.
Learning by playing. His method inverts the usual advice to learn all the math, then programming, then machine learning. Hand the child the ball first. Pick a dataset tied to a hobby (his was tennis), plot it with just the line of code you need, and learn linear regression when a line appears. He applies the same approach to learning piano by songs before chords. As a demonstration he explains gradient descent as getting down a foggy mountain one locally steepest step at a time, with the loss function as your height.
What machines will not take. Serrano lays out a pyramid. Physical strength was displaced by machines, rational thinking is being displaced now (chess for more than 20 years, go, and in his prediction soon math and physics), and above both sit emotional intelligence and intuition, which he says Western society has neglected and calls "soft skills" when they are the hard ones. He expects future humans to seem less capable to us, as we would to our ancestors, while holding skills we cannot imagine. His heuristic is the 1% of human input left after automation, like telling a car where to go. In his own work, the part of making videos with language models that stays his is the questioning: he steers the model as a child would with repeated whys. The hosts connect this to executives, who also mostly ask questions.
Quantum computing. He explains qubits as switches that can be on and off at once, and entanglement as stranger than a pair of gloves split across galaxies. The practical gain is on problems that grow exponentially, such as the traveling salesman problem and factoring, which would break RSA. He expects quantum hardware to be a small step inside mostly classical computation, like a bus that only sometimes flies. He believes, flagging that he may be wrong, that quantum machines are better at generative AI because they generate true randomness, citing his own small experiments. He goes on to speculate that a quantum measurement's randomness might be where empathy or intuition, or even consciousness, could enter a machine.
Underrated skill. Asked the show's closing question, he names malleability and says knowledge is overrated now that it no longer lasts 50 years. When hiring, he weighs how many new things a person has learned, and compares that to a marathon runner who will manage an Ironman before learning to swim.
- Most people are taught math in a language they don't speak. Luis Serrano of Serrano Academy argues that abstraction is the native tongue of a small academic minority, and that concrete framings unlock what people already know: few can draw a written equation, but everyone draws a parabola for a tennis ball falling off a building. His own case is that he was failing grade-8 math in Colombia, then placed first in the country on a puzzle-style olympiad exam.
- Learn machine learning by playing first, theory later. His analogy is that you hand a child the soccer ball before the rulebook. He advises taking a dataset from a hobby (he used tennis), plotting it with only the line of code needed, and learning linear regression when a line shows up. He demonstrates the method by explaining gradient descent as walking down a foggy mountain, with the loss as your height.
- Make AI do your whole job to find what's still yours. His test is to try to get the computer to do 100% of the work and see the 1% where it falls short. Serrano's pyramid runs from physical to rational to emotional and intuitive intelligence. Machines already beat humans at chess and go, and he predicts math and physics come next, so he expects the unreplaceable part to sit higher up. He makes his videos with language models, and the part that is still his is the questions he asks; asking a question, he says, is harder than answering one.
- Quantum computing will be a small quantum step inside classical machines. His analogy is an airplane that is a bus most of the time and only uses its wings for the step that would otherwise take a million years, such as factoring, which would break RSA. He believes, hedging that he may be wrong, that quantum machines suit generative AI and classical ones suit supervised learning. In his own small experiments, quantum generators produced output further from the training data. He then speculates that he cannot rule out quantum randomness carrying empathy or intuition.
- Malleability matters more than knowledge for data professionals. This is his answer to the show's most-underrated-skill question. He says knowledge is overrated because it no longer lasts 50 years, and when hiring he looks at how many new things a candidate has learned. His analogy is that someone who runs marathons will manage an Ironman even before learning to swim.
Crosstalk SolutionsInside Ubiquiti's Booth at RSAC 2025 – New Gear Revealed!
- 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.
Polis Pandit$A New YouTube Strategy
- This post includes paid content that is not available publicly. See the original source.
Scott HanselmanEPISODE 15 - Scott & Mark Learn To… Vibe coding and its limits (Part 2)
synopsis
Looking back on the fractal shader they vibe coded for Windows Terminal in the previous episode, Mark ends up arguing that vibe coding suits hobby prototypes. In his view, context limits inherent in transformer models, brownfield maintenance and enterprise accountability mean people who know what they are doing must still own the code.
The one-shot result. The shader took under 15 minutes. Hanselman puts the hand-written version at four to six hours, and the copy-from-42-browser-tabs version, which he calls "Franken coding" and they settle on calling "Stack Overflow coding", at two to three. Mark had given the attempt about an 80% chance with an example file supplied. The first try, on a thinking model, failed. Mark says that because the output is non-deterministic, the success may have been a lucky draw.
Why context doesn't scale. Mark cites the needle-in-a-haystack tests in Microsoft's Eureka benchmarks, which frontier models can fail at 32k or even 8k tokens. He also cites the "Lost in the Middle" paper. A whole enterprise app pasted into a million-token window will not be reasoned over reliably.
From greenfield to brownfield. A vibe-coded v1 instantly becomes an inherited codebase. Mark describes models losing state, dropping functionality and claiming fixes they did not make, all in a loop of "Oh, I see what happened". Hanselman says nobody cares that the code came from AI when feature X is due.
Enterprise accountability. Mark lists the risks: unvetted packages, deprecated API versions caused by training cutoffs, and specs that are only about 80% written. Hanselman adds the unwritten institutional knowledge an enterprise carries.
- 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.
Dr. Jon KrohnInside NVIDIA GTC: The Super Bowl of AI Conferences
- 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.
Technology and FriendsPeter Van Vliet on Separating the Development Model From the Deployment Model [EPISODE 851]
- 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.
Dr. Jon Krohn883: Blackwell GPUs Are Now Available at Your Desk — with Sama Bali and Logan Lawler
synopsis
Sama Bali, who leads AI solutions go-to-market at Nvidia, and Logan Lawler of Dell's Pro Max workstation line argue that the economics of AI development have swung back from the cloud to the desk: data-centre GPUs are scarce, experimentation needs a private sandbox, and Blackwell-based workstations, paired with Nvidia's software stack, now carry enough memory and compute to fine-tune and serve large models locally. Their endpoint is a pair of Linux systems, the palm-sized GB10 and the tower-sized GB300, pitched as server power for teams that have no servers.
The rebrand and the launch. Lawler explains that Dell collapsed its many PC brands into Dell, Dell Pro and Dell Pro Max, the last for heavy GPU-bound workloads. The line was announced at CES and launched at GTC, about a week before recording. Beyond the GPU, he notes the Pro Max T2 tower gains room for drives and cards and runs new Intel processors at a 250-watt sustained load.
Why local hardware again. Bali says Nvidia saw demand from people fine-tuning and running models themselves, and recounts a data scientist who queued in a Slack channel to get a cloud instance. The RTX Pro Blackwell cards answer this with memory doubled from 48 to 96 GB per GPU, four to a workstation. Lawler offers a rule of thumb of about 2 GB per billion parameters, varying with precision, and says models such as Llama 3 405B were previously out of reach on a workstation.
The software layer. Bali walks through NVIDIA AI Enterprise. NIM microservices containerise open-source, proprietary and Nvidia models with inference optimisations already applied, so they need no manual tuning to the GPU and can be swapped as model versions turn over every couple of months; Lawler's practical definition is one line of code to stand up a model. She points to free prototyping on build.nvidia.com and free downloads through the developer programme, then NeMo for training, fine-tuning and guardrails, AI Blueprints as reference workflows, and CUDA libraries: RAPIDS cuDF accelerating pandas-style preprocessing by up to 100x without code changes, and cuML, announced at GTC, up to 50x on scikit-learn tasks. Lawler says the same software is what lets work move from a Pro Max to an in-house server for deployment.
Linux or Windows. Lawler draws a line: the GB10 and GB300 come only with Linux and the Nvidia stack preloaded, so they suit people doing nothing but data science, while anyone needing applications such as Adobe Creative Cloud should buy a conventional Windows Pro Max. Bali says users need not choose, since a GB10 can sit beside a Windows laptop, and two can be daisy-chained over ConnectX in less space than a typical tower.
The GB300 and cost. Lawler gives the GB300 784 GB of unified memory, 288 GB of it GPU memory, and 20,000 TOPS at FP4, against about 4,000 for a single RTX Pro 6000. Bali expects several team members to share one; Lawler pitches it at mid-market firms without server skills, racks or cooling. Pricing was unreleased at recording. Lawler guesses the GB10 will land near Nvidia's DGX Spark, which he puts at "3 to 4,000", and says the GB300 will cost more but not as much as a server.
Where it goes next. Lawler predicts AI reaching non-developers like himself, who trained LoRAs for an animation studio, and everyday uses such as a RAG system over his mother's recipe cards. Bali, following Jensen Huang's keynote, describes a move from generative to agentic AI, with agents coordinating supply, production and transport in a factory, and then physical AI such as driverless cars, while hoping for guardrails and regulation.
- 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.
Crosstalk SolutionsFirst Look! Come see the UniFi Switch Pro XG Lineup
- 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 $499 Pro XG8 PoE is, as far as the host knows, the first desktop switch with Etherlighting.
- Ubiquiti uses its own name for 90 W PoE. The host explains 802.3bt Type 3, known as PoE++, supplies up to 51 W, while Type 4 at up to 90 W carries an extra-plus name he believes only Ubiquiti uses; the only device he knows that accepts it is the Switch Flex 2.5G PoE.
- These are among UniFi's cheapest switches for powering 10G access points. The host names the Enterprise E7, U7 Pro and U7 Pro XGS as targets; the Pro XG24 offers 720 W of PoE and 460 Gbps switching capacity, drawing 150 W bare and up to 870 W fully loaded.
Polis Pandit$Crossing the Rubicon By Arresting Judges
- This post includes paid content that is not available publicly. See the original source.
Dr. Jon KrohnWhat’s Next for AI? Emily Webber on the Future of LLMs & Hardware
- 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.
Dr. Jon KrohnTrainium vs. Inferentia: How to Choose the Right AI Chip (with Emily Webber)
- 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.
Dr. Jon KrohnWhy Use Trainium Instead of a GPU? Emily Webber Explains
- 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.