Category Archives: AI

Language Models for Text Classification: From Bag-of-Words to Jev

Posted on by
Reply

Software engineering builds on previous work, and the same goes for LLMs. Having worked with both enough to recognize the parallels, I find it interesting how many problems we now approach with general-purpose LLMs have roots in methods we have been using for years.

Sebastian Raschka’s piece on text classification is a useful reminder of that progression: bag-of-words representations with logistic regression or Naive Bayes, recurrent and convolutional networks, transformers, and now models positioned between specialized classifiers and general-purpose LLMs. The engineering question is therefore how much capability the workload actually requires.

For a bounded classification problem, model selection affects latency, inference cost, operational complexity and the amount of nondeterminism introduced into the system. A general-purpose LLM provides flexibility across tasks, while a specialized classifier can offer a narrower execution path when the input space and expected outputs are well defined.

My time with software engineering and data science has made this trade-off particularly familiar to me. The underlying principle has not changed much: understand the problem, define the constraints, then choose the least complex approach that satisfies them. LLMs expand the set of tools available to us, but they do not remove the architectural decisions that come before choosing one.

https://magazine.sebastianraschka.com/p/classifier-history-and-jev

Introducing Strand Harness

Posted on by
Reply

A 28% reduction in token cost without changing the underlying model immediately caught my attention, mostly because the optimization is happening somewhere that receives less attention than model selection: the agent harness.

The Strands team published benchmark results for its new agent harness, reporting 28% lower cost when using the same Claude or GPT models across six benchmarks, while maintaining similar accuracy. The implementation details are more interesting than the headline number: tool results above roughly 1,500 tokens are truncated, compaction starts after context usage exceeds 85%, and context recovery is handled inside the execution loop. Those mechanisms show how much of agent performance is now an orchestration problem rather than a model selection problem.

This feels particularly relevant with the attention around Jev from TypeSafe AI, which approaches the same problem from another direction by moving structured decisions away from general-purpose LLM generation altogether. We are starting to see the architecture around models become as important as the models themselves. The question is shifting from “which model should we use?” toward “which parts of the system actually need a general-purpose model, and how should intelligence be orchestrated around them?” Context lifecycle, model routing, typed decisions, tool boundaries, failure recovery and token economics are increasingly becoming architectural decisions in their own right.

#AI #AgenticAI #SoftwareArchitecture

https://strandsagents.com/blog/introducing-strands-harness/

Extensible Software in the age of LLMs

Posted on by

Most software products are built for the top of a demand curve, where a limited set of features serves the largest group of users, while a long tail of individual needs is left unaddressed because supporting it does not scale in the traditional sense Most software products are built for the top of a demand curve, where a limited set of features serves the largest group of users, while a long tail of individual needs is left unaddressed because supporting it does not scale in the traditional sense (and attempting to do so often leads to scope creep)..

Jeremy Morrell argues that large language models change this calculation, since they lower the cost of authoring an extension to the point where serving a single user’s edge case becomes economically viable, provided the surrounding platform can deploy and secure that extension safely.

The interesting part of the argument is not that LLMs can generate code, which is now a familiar claim, but that the harder problem has always been the deployment and trust boundary around that code, not its authorship. Webhooks require the user to operate a separate service and handle delivery failures themselves, and most local extension models, such as IDE plugins or Obsidian’s ecosystem, ask the user to trust every author whose extension they install, which is a workable tradeoff for a notes app but not for a system holding financial transactions or private messages. Salesforce solved a version of this problem two decades ago by building its own compiler, runtime, and standard library so that customer logic could run safely inside a shared multi-tenant platform, well before serverless made that pattern common.

What makes the current moment diffferent is that the primitives Salesforce had to build from scratch now exist off the shelf, whether as V8 isolates, WebAssembly with WASI, or microVMs, and each comes with a different tradeoff between cold start latency, isolation strength, and language flexibility. Morrell’s argument for an object capability model over a traditional API proxy is worth sitting with regardless of which runtime a team chooses, since restricting untrusted code to explicit, narrow references such as getApprovedEmail() rather than a scoped API token removes an entire category of exfiltration risk by construction, instead of relying on a proxy layer to catch every malicious pattern after the fact.

For teams building internal tooling or customer facing platforms, the practical takeaway is that extensibility is no longer purely a UX or roadmap decision, it is also an infrastructure decision, and the primitive chosen early on will determine how much custom logic a platform can safely absorb once users start asking an LLM to write it for them.

https://jeremymorrell.dev/blog/extensible-software-in-the-age-of-llms/

Watermark

Posted on by

It took only days to have an easy way to remove the watermarks. There are already ~70 repos:

https://github.com/search?q=watermark+remover+created%3A%3E2026-08-04&type=repositories

Posted in AI

World Bank: The Promise of Artificial Intelligence

Posted on by

Elon Musk knows how to promote himself, he said recently that AI and robots will replace all jobs, and that working will become optional. It is a memorable line, and it travels fast. I read last week the World Bank’s World Development Report 2026: The Promise of Artificial Intelligence, and the picture it paints is far more grounded. It is the first comprehensive assessment of what AI means for the 6.8 billion people living in low and middle income countries, and the jobs finding deserves attention. Roughly 4.5 percent of jobs in those economies are amenable to automation by generative AI, against 14.2 percent in high income countries, while 16.2 percent are amenable to being complemented by AI. The report is direct about it: AI is not yet replacing large numbers of jobs in most developing countries. For most of the world, the immediate story is augmentation, not elimination.

I am an optimistic guy but not to the level of Pollyanna. The report is candid about real pressure on call centres and business process outsourcing, about early evidence that entry level white collar postings have thinned in India and China (i see this in my banking industry for sure), and about the dependency risk that comes with a handful of companies controlling the frontier. But the binding constraint it identifies is not runaway machines. It is reliable electricity, internet access, education, and local language data. In Sub-Saharan Africa, nearly one third of rural schools still lack dependable power. It argues that countries should be blinded neither by the hype nor the hysteria, and offer a plain framework instead: adopt, adapt, then advance. What stays with me is how much of this looks less like a technology story and more like an old fashioned development story wearing new clothes.

https://www.worldbank.org/en/publication/wdr2026

The collapse of the web as we know it

Posted on by

On March 5th, TheNumbers.com, the film industry’s most trusted box office database, went dark without warning. It came back a week later stripped down, missing historical charts, movie pages, and its report builder. The cause was not a single dramatic event, it was a slow collapse under a combination of pressures that most small, independent websites are not built to survive. Founder Bruce Nash later revealed that only 10% of the site’s traffic came from actual human visitors, the rest was AI crawlers and agentic bots hitting a thirty year old system with 160,000 legacy files. Buried in that traffic were also signs of deliberate probing, likely aimed at accessing box office data before it went public, since prediction markets use those numbers to settle real money bets.

The deeper issue here is not that one site got hacked, it is that the assumptions the open web was built on no longer hold. Data quality and traffic volume are no longer reliable indicators of a site’s health, resilience against automated extraction now matters just as much, and any business relying on a single, aging web presence should treat that infrastructure as a liability to be actively managed, not a static asset to be left alone. And scarier, for the small website mantainer, it is a Davi-Goliath battle that most of them does not know how to fight.

https://stephenfollows.com/p/what-just-happened-to-thenumberscom-should-worry-us-all

Entire.io

Posted on by

I just got my invire for entire.io, it is really cool, incredibly ui, easy cli. It does not track only what changed in the commit, it captures the entire session I had with claude, the prompt, answers, tool calls, token usage, top! I was mantaining a very naive tracker keeping the guids so I could return to a session, dont need it anymore. well, until they charge me an arm and a leg 🙂

https://github.com/sergiorgiraldo/entire101/

Learn to Code

Posted on by

I am using ChatGPT since dec/22 and I showed to my son in that same month, I can say the boy turned into a llm master. Fast forward to aug/25, he asked if he should learn Python during his gap year and I said “for sure!” with all the enthusiasm to see your son doing smth you love 🙂 but …I know there is a growing assumption in technical circles that learning to code has lost its purpose, now that language models can generate working software from a short prompt, like my kid could do in a blink.

I stand by my suggestion and I read an article I loved about this clash of opinions. The argument treats coding the way we already treat mathematics or literature, as a discipline worth learning for what it teaches regardless of direct vocational payoff. The skills gained through the learning process extend well beyond syntax. Debugging teaches a structured way of isolating the source of a problem; composition teaches how small, well defined pieces combine into something larger; and the discipline of unambiguous instruction transfers to almost any field that requires clear thinking. These are meta-skills, in the sense that they remain useful long after any specific language or framework becomes obsolete. And with the added bonus of knowing to program 🙂

I could not have said better!

https://stevekrouse.com/learn-to-code

What happened after 2,000 people tried to hack my AI assistant

Posted on by

The developer Fernando Irarrázaval ran a public experiment where anyone could email his AI assistant and try to make it leak the contents of a secrets file, and the results clarify where prompt injection stands today. Over the course of the experiment the assistant received more than six thousand emails from over two thousand people, and not one of them succeeded in extracting the secret or triggering an unauthorized reply.The defensive setup was minimal; the system prompt contained only a few lines instructing the model never to reveal credentials, never to modify its own files, and never to execute code received over email.

The attacks covered the range of social engineering you would expect, including authority impersonation, fabricated incident response requests, fake compliance audits, and the same message rewritten across several languages to probe for weaker instruction-following outside English. And his defenses were effective to the point of non-exploitation.

I’ll push back a little. Given how consistently security researchers flag prompt injection as a real, unresolved risk for agentic systems, I don’t think one experiment, even if well-run and designed, should move anyone toward optimism. It shows a hardened model resisted attacks over email. It doesn’t show the problem is smaller than experts think.

https://www.fernandoi.cl/posts/hackmyclaw/