Issue 74 — 2026-08-05
DeepSeek drops V4-Flash, a new open AI model on Hugging Face
DeepSeek released DeepSeek-V4-Flash-0731 on Hugging Face, a July 31 model in their Flash line. The model is available under an open license, which means developers, researchers, and anyone with a GPU can pull it down and run it locally or on their own infrastructure. No API keys, no vendor lock-in.
Flash models from DeepSeek have been getting attention for punching above their weight — good performance relative to size, which matters when you're running on consumer hardware or tight compute budgets. The July 31 version updates the Flash family, though the exact benchmarks and improvements aren't spelled out in the release notes.
Why this matters for us: open models let Brown and Black founders, indie devs, and community orgs build without paying Big Tech's toll — no gatekeepers, no rate limits, no terms of service that can change overnight.
Microsoft's new ML compiler speeds up models across CPUs and GPUs
Microsoft released MSLK, a compiler that takes machine learning models and rewrites them so they run faster on whatever hardware you have — CPUs, GPUs, you name it. The idea is straightforward: you train a model once, then MSLK figures out how to make it execute efficiently on your machine. No more rewriting the same model five different ways for different backends.
What makes this interesting is that ML workloads are everywhere now — from recommendation engines to the AI features getting baked into everyday apps — and every time you deploy a model, you're making hardware choices that affect cost and speed. MSLK sits between the model and the chip and optimizes the execution path. If you're running inference at scale, this is the kind of tool that quietly saves money on every request.
Why this matters for us: the faster and cheaper these models run, the more likely they end up in the tools we use every day, and the less we pay for the compute that powers them.
Qué es el system prompt — y por qué le importa a la migra app
Todo modelo de IA tiene dos partes. Una es lo que le dices — el user prompt. La otra es el system prompt, las instrucciones que nadie ve pero que le dicen al modelo cómo comportarse antes de que le preguntes algo.
Imagina a tu primo que siempre está listo para ayudarte. Si lo mandas a una fiesta sin decirle qué hacer, te va a dar un desorden. Si le dices "sólo habla en español, sé directo, no inventes datos," y listo, ya sabes cómo va a responder.
El system prompt es esa primera instrucción. Las empresas lo usan para darle personalidad al chatbot, fijar límites, o darle contexto de su trabajo. Un bot de banco puede llevar un system prompt que le dice: "no des información médica, si el usuario pregunta sobre recetas, redirige a un profesional." Ese mensaje está ahí antes de que tú hables.
¿Por qué te importa?
Porque la mayoría de los bots que usas — el de tu banco, el de la app de la migra, el de la escuela — tienen un system prompt que define qué puede y qué no puede hacer. Y si quieres que un bot funcione bien para tu negocio, lo que escribes en ese primer mensaje define todo el resto.
Tip: Si estás probando un bot para tu negocio, abre el chat y hazle una pregunta que esté al borde de lo que debería responder. La forma en que se porta es el system prompt hablando por ti.
quality has to make money or it doesn't exist
— benn.substack.com
#quality-has-to-make-money-or-it-doesn-t-exist-b441f0Apple finally lets iPhone copy-paste into Windows
Apple shipped a feature that lets you copy text on your iPhone and paste it directly into Windows — no shared folders, no Bluetooth tricks, no jumping through hoops. It's been a real pain for years. People who use both platforms have been working around it for as long as I can remember, and now Apple is finally shipping it.
The feature works when both devices are signed into the same Apple ID and connected to the same Wi-Fi network. Text, URLs, notes, and links all flow across. It's not fancy, but it's the kind of thing that saves time in a way that matters for people who actually do the work.
Bilingual AI that handles your docs — and your data
Your school district, community clinic, or county program needs AI that actually works in the language your people use. Not one that drops Spanish mid-conversation. Not one that sends sensitive patient records to a public cloud.
BFTS Chat gives your org its own tenant with eight purpose-built tools: chat, doc analysis, IEP drafts, helpdesk, grants, prior auth, SOPs, and proposals. Bilingual by default. Spanish and English in the same thread, no switching required.
You can run it against an on-prem brain instead of a public API. Your documents stay where they belong.
Why this matters for us:
La gente deserves tools that handle their language without risking their data or their families' information.
https://tools.brownforces.io
Anthropic investigates AI-powered cyber incidents as Google chases Gemini on desktop
Anthropic just released a report on how it handles AI-powered cyber incidents — the kind of threats that show up in primos' WhatsApp groups, the kind that make la gente lose money. They're tracking how bad actors are using AI models to craft phishing, automate attacks, and probe systems. The company's also publishing its cybersecurity evals, so other teams can see what they're actually testing against.
Meanwhile, Google's trying to close the feature gap on Gemini for desktop. It's the kind of catch-up game we've seen before, where one company ships first, another ships later and claims they did it better.
Qwen 3.8 drops — more open-weight firepower from China's AI camp
Qwen 3.8 has landed. The latest model in Alibaba's Qwen line is an open-weight AI model, meaning the weights are publicly available rather than locked behind a paywall. Open-weight models let developers, researchers, and small teams run the model themselves without depending on a single vendor's API.
Qwen's been building a serious open-source presence. The broader Qwen family has become one of the more popular alternatives to the Big Three in the U.S. AI space, and this release adds another rung to that ladder.
Why this matters for us:
Canva's AI push: one percent of the way there
Canva, the design tool built for regular people (not just designers), is still figuring out where AI takes them. The company's been building out AI features — image generation, text-to-design, smart editing — and the verdict from their own assessment is that they've barely scratched the surface of what this could become.
That's actually the right way to think about it. The tools that let anyone make a poster, a flyer, a social post, or a pitch deck are still in the early innings. Canva's core users — small business owners, teachers, community organizers, abuelas running a home-based food business — aren't asking for a full Adobe.
GitHub launches gh-stack for building coding agents
GitHub just released gh-stack, a framework for building coding agents — AI tools that don't just write code, but actually run it, debug it, and iterate. The stack bundles three pieces into one architecture: a context window that keeps track of what the agent has seen, a tool integrations layer so it can talk to your repos and terminals, and an execution engine that lets the agent do real work instead of just generating text.
The idea is straightforward. Right now, a lot of coding assistants are really fancy autocomplete. gh-stack is trying to push past that by giving agents the full loop — read, write, run, debug, repeat. It's open source, it's designed for developers who want to build their own agents rather than rely on someone else's, and it ships alongside GitHub's existing Copilot tools.
Why this matters for us: more agents mean more tools for the people who actually write code day to day — and the ones who can build their own end up with a real edge over the ones still waiting on the next SaaS subscription.
GM and Ford Are Dialing Back on EVs
GM and Ford are mentioning electric vehicles on their quarterly investor calls at pre-pandemic rates, according to new data from TechCrunch and Hudson Labs. That's a sharp drop from the peak years when EV talk dominated earnings calls, strategy decks, and CEO speeches. The industry that spent billions betting on electrification is now talking about it like it's just another product line.
What's behind the silence is a familiar pattern. EV sales growth has slowed across the board, charging infrastructure has lagged behind promises, and consumers are still working out whether the math makes sense when gas prices fluctuate.
Spotify's engineering post on indexing data lakes for fast queries
Spotify's engineering blog published a post on indexing data lakes so you can do fast point queries — the kind of thing that normally requires a separate database layer. The gist: you can build on top of cheap storage and still get sub-second lookups if you index right.
It's part of the broader shift toward lakehouse architectures — ditching the expensive data warehouses and trying to run everything on S3-like storage with smart indexing. The trade-off is real: cheaper storage, but you have to solve the query performance problem yourself instead of outsourcing it to a managed service.
Why this matters for us: if your team's drowning in data costs or slow queries, this is a concrete angle worth looking at — and it's a reminder that the tools we rely on are built by engineers who've been through the same pain.