sgidi

PORTAL SGIDI

Sistema de Gestão da Investigação, Desenvolvimento e Inovação 

• Repositório: Google Drive: pasta NI&D/SGIDI (link)

• Repositório Github: link

• Gravações das reuniões NID (Zoom): aqui 

• Norma  Portuguesa NP 4457 2021 (pdf)

• Norma Portuguesa NP 19011 (2019) (pdf)

• MG 01_00: Manual SGIDI (pdf) (versão controlada: pdf)

• MG 02_00: Manual de funções (doc)

• RG 07_01: Plano de auditoria interna (junho 2023, pdf)

• RG 08_01: Relatório da auditoria interna (junho 2023, doc) (auditor: cv0, cv1, cv2)

• RG 07:02: Plano de auditoria externa, fase 1 (julho 2023, pdf)

• RG 08_02: Relatório de auditoria externa, fase I (julho 2023, pdf)

• RG 07_03: Plano de auditoria externa, fase 2 (setembro 2023, pdf)

• RG 08_03: Relatório da auditoria (fase II) de concessão (setembro 2023, pdf); Certificados (pdf): pt, en, es

• RG 08_04: Relatório da auditoria de acompanhamento (setembro 2024, pdf)

• RG 07_04: Plano de auditoria interna (dezembro 2024, pdf)

• RG 08_05: Relatório da auditoria interna (dezembro 2024, pdf) (auditor: cv0, cv1, cv2)

• Mapas de processos: 

  • PR 01_00: Controlo da informação documentada (doc)

  • PR 02_00: Controlo da conceção (doc)

  • PR 03_00: Gestão de ideias e iniciativas de inovação (doc)

  • PR 04_00: Gestão das interfaces (doc) 

  • PR 05_00: Recursos humanos (doc)

  • PR 06_00: Auditorias interna (doc)

  • PR 07_00: Não conformidades, ações corretivas e de melhoria (doc)

• Registos: 

  • RG 01_00 Matriz de uso e controlo de documentos (xls)

  • RG 04_01 Matriz de de ideias e projetos (xls)

  • RG 05_00 Fichas de ideias e projetos (aqui) e respetivos cronogramas (aqui, botão)

  • RG 06_00 Ocorrências e planos corretivos (doc) (xls) 

  • RG 07_00 Plano de auditoria (doc)

  • RG 08_00 Relatório de auditoria (doc)

  • RG 09_01: Matriz de riscos e oportunidades (doc)

• Planos: 

  • PL 01_02: Plano de formação 2025 (doc)

  • PL 01_01: Plano de formação 2024 (doc)

  • PL 01_00: Plano de formação 2023 (doc)

    • • Ação de formação: Gestão da Investigação, Desenvolvimento e Inovação - NP 4457:2021 (Iberogestão,  11 abril 2023) (pdf). Certificados (pdf).

      • Ação de formação: Auditorias a sistemas de gestão ISO 19011:2018 (Iberogestão, 10 julho 2023) (pdf). Certificados (pdf).

  • PL 02_00: Orçamento IDI (xls)

Atas de acompanhamento e revisão pela gestão:

• Ata #12, 21 de julho de 2025, revisão pela gestão (doc) 

• Ata #11, 21 de dezembro de 2024, revisão pela gestão (doc)

• Ata #10, 2 de julho de 2024, revisão pela gestão (doc)

• Ata #9c, 31 de março de 2024 (doc)

• Ata #9, 22 de dezembro de 2023, revisão pela gestão (doc)

• Ata #8b, 31 de outubro de 2023 (doc)

• Ata #8, 1 de Agosto de 2023 (doc)

• Ata #7, 1 de Julho de 2023 (doc)

• Ata #6, 23 de junho de 2023, revisão pela gestão (doc)

• Ata #5, 1 de Junho de 2023 (doc)

• Ata #4, 1 de Maio de 2023 (doc)

• Ata #3, 1 de Abril de 2023 (doc)

• Ata #2, 1 de março de 2023 (doc)

• Ata #1, 1 fevereiro de 2023 (doc)

Documentos preparatórios:

• Registos online: workflow e sistema documental (doc)

• Contexto externo e fontes para vigilância tecnológica (doc)

•  [v11 (doc) ]

ID25. 20 de janeiro de 2025, EB
ia.arq: consorcio e plataforma,  ID25, PJ13
Organizar um consorcio de empresas e entidades académicas para apresentar uma candidatura aos fundos (SIID empresarial em copromoção) para desenvolver e testar uma plataforma na cloud para arquivos documentais e jornais com uma nova arquitetura IA e recorrendo a agentes de IA para funcionalidades avançadas de procura e investigação.

ID23. 3 de setembro de 2024, EB PROJETO CONCLUIDO
Workflows "low cost" de arquivo e exploração de arquivos documentais históricos, II23,PJ11
As facilidades de arquivo de documentos digitalizados e as facilidades de procura em serviços como o Google Drive e IA abrem novas perpetivas para a investigação em dominios como a historiografia ou em coleções de história natural. Ferramentas adicionais  faceis de integrar no ambiente do Google Drive facilitam ainda mais o trabalho de recolha de informação e a transcrição de textos.

ID21. 20 de janeiro de 2024, EB
Geração de dados em formato gbif em fichas de herbário
Desenvolver app com facilidades de geração de quadros em Excel e etiquetas em formato padrão gbif a partir das fichas de herbário transcritas pela plataforma (dados e metadados)

  https://www.gbif.org/standards 

ID19. 15 de dezembro de 2023, EB
Integrar facilidades de reconstrução de textos de noticias de jornais via ChatGPT na plataforma de visualização e procura de jornais antigos, II19,PJ9 PROJETO CONCLUIDO
Expandir os resultados obtidos em PJ6 para integrar de forma interativa as facilidades de reconstrução de textos selecionados em páginas de jornais usando os recursos de chatGPT

31 de Março de 2023, EB

Modelos fundacionais e tradução

O GPT & Ca. é considerado "uma maravilha de multilinguismo" (ultimo The Economist) e uma ferramenta muito capaz de tradução, especialmente em linguas "com mais recursos". Parece ser consensual que o futuro da tradução será hibrido com "pós processamento editorial" humano e que a tradução trivial será feita por AI (LLM combinado com outros modelos MT). Ler os jornais antigos é, ao fim e ao cabo, um problema de tradução: passar do texto com erros para o texto sem erros e eventualmente convertido para a linguagem (estilo) atual. As experiencias de "tradução" que fizemos (reescrita de partes de textos de noticias originais) parecem confirmar isto.
Logo uma app que use um modelo LLM refinado para o contexto dos jornais (o universo linguistico de base) deverá permitir consultar / traduzir o texto dos jornais e responder a perguntas (prompts) dentro desse universo. Isto significa passar por cima da questão do reconhecimento correto dos carateres gráficos mesmo correndo o risco de alguns erros. Parece semelhante a traduzir crioulo ou mirandes para portugues corrente ou vice versa. Esta abordagem à leitura e "search" nos jornais ultrapassarria as dificuldades do reconhecimento de carateres e permite "salvar" scannings já feitos e com eventual com pior qualidade (caso dos nossos jornais).
O q compreendi do VertexAI e do GenAppBuilder da Google é que permitem fazer isto com facilidade e rapidez e são o "frontend" de abertura da tecnologia pela Google para inserção de facilidades AI nas apps de processos correntes. O "tuning" (ou "fine tuning") do Vertex parece-me ser isso, um processo de ML machine learning automatico que compreendo corresponder a restringir o universo de aplicação do LLM ao contexto proporcionado pelos dados (neste caso, os textos txt das colunas dos jornais). Presumo que "tokens" que não aparecem nos dados não devem ser usadas (seja o que for q se entenda por token neste caso) pela app: parece corresponder a gerar um modelo LLM simplificado para os dados submetidos. Este treino do modelo parece ser feito automáticamente pelo Vertex AI com a possibilidade de escolha de diferentes modelos fundacionais (o "model garden").

• • Potencial do Vertex AI como "tradutor" de ocr com erros (jornais antigos)

  • Comentários: aqui, aqui

ID16. 10 de março de 2023, EB

Experiencias VR/MR em museologia digital com hololens

• Potencialidades da VR/MR com holow lens (Microsoft) para museologia digital

• aqui e aqui

ID14. 29 de Fevereiro de 2023, EB

Expansão da Inovatec Press para os mercados de lingua portuguesa

• Integrar produtos da Inovatec Press em plataformas brasileiras com print on demand e plataformas nacionais com distribuição digital para os paises de lingua portuguesa

ID12. 28 de Fevereiro de 2023, JB

Potencial de tecnologias audio 3D para performances e instalações Datagrama

• Potencial das tecnologias audio ambisonics 3D em performances e instalações artisticas Datagrama (camada adicional nos modelos Datagrama de VR/MR: speakers com geometrias espaciais de alinhamento,  complementaridade sensorial)

• https://youtu.be/51dUnndYwo4 

ID11. 13 de Fevereiro de 2023, JB; II11

memTSI:UP Memória das tecnologias de informação na Universidade do Porto

• Uma década depois, a experiencia adquirida com o projeto memTSI pode ser interessante para uma nova iniciativa desse tipo, mas agora restrita ao universo da Universidade do Porto e ao papel da UP na inovação tecnológica neste domínio ao longo de um período histórico de três décadas e do seu impacto territorial e nacional. Muitos dos primeiros atores desta história, desde os anos 60, estão ainda vivos, pelo que uma iniciativa deste tipo parece oportuna e segue as tendências internacionais para preservar a memória das tecnologias da informação com base em história oral (entrevistas) dos protagonistas

ID9. 20 de novembro de 2022, NB; II9

Emidio Biel e a linha do Tua

• As fotografias de Emilio Biel sobre a linha do Tua continuam pouco conhecidas. Uma exposição dessas fotos com um programa associado de sessões sobre E. Biel e sobre a a história da linha do Tua pode constituir uma oportunidade com a Universidade do Porto (Galeria da Biodiversidade) e para um "road show" pelos cinco concelhos do vale do Tua (em colaboração com a Agencia Regional do Vale do Tua) e assim difundir os trabalhos desenvolvidos no projeto FOZTUA sobre o assunto (Cf. livro "A linha do Tua, 1887, e as fotografias de E. Biel"

ID7. 31 de julho de 2022, EB; II7

Digitalização de manuscritos e fichas da Sociedade de Geografia de Lisboa

• A coleção de manuscritos da SGL justifica a sua digitalização urgente. A vasta biblioteca não está disponivel online para consulta. Os testes feitos em visita recente sugerem a viabilidade da iniciativa e o interesse institucional. As plataformas online desenvolvidas até aqui podem ser fácilmente aplicadas neste caso

ID5. 1 de junho de 2022, EB; II5, PJ4 PROJETO CONCLUIDO

Digitalização de fichas de coleções de história natural

• Muitas coleções de história natural estão catalogadas em fichas manuais (geralmente manuscritas, por vezes datilografadas) cuja digitação se tem mostrado inviável por falta de recursos. A experiência com as fichas bibliográficas do Ateneu Comercial do Porto sugere a sua extensão às fichas de coleções do Museu de História Natural e Ciência da Universidade do Porto e a extensão da plataforma na cloud desenvolvida para as fichas bibliográficas do Ateneu para este tipo de universo.

ID3. 10 de novembro de 2021, EB; II3, PJ2 PROJETO CONCLUIDO; INSTALAÇÃO 

Digitalização de fichas bibliográficas manuscritas

• O catálogo da biblioteca do Ateneu Comercial do Porto (mais de 50 mil livros) em fichas manuscritas pode ser um bom teste para o uso dos scanners de nova geração (CZUR para documentos planos e Epson ES para fotografias / fichas). Uma extensão da plataforma cloud desenvolvida para jornais (séc. XIX) poderá ser adaptada para permitir a procura e acesso às fichas bibliográficas sem necessidade da sua digitação manual.

ID1. 1 de dezembro de 2018, EB; II1

memDÃO: Memória da linha do Dão

•  Ao longo de quase um século as linhas do Dão (entre Santa Comba Dão e Viseu) e a linha do Vouga (entre Espinho e Viseu) foram importantes vias de transporte de passageiros e carga, que ajudaram ao desenvolvimento económico e social da cidade e contribuíram para a atual configuração da região. Muita da experiência humana de região foi influenciada pelas linhas do Dão e do Vouga, sendo importante recolher, preservar e valorizar a memória dessas vivências. Com base na experiência do projeto FOZTUA propõe-se um projeto para produzir materiais sobre a história e património ferroviário da região que valorizem infraestruturas regionais (ecopista do Dão, por exemplo). 

1 de outubro de 2025, EB

https://mail.google.com/mail/u/0/#inbox/FMfcgzQcpwwLnCmMxwkzMCtSjHhgDNNt

Ex-OpenAI CTO Makes a $12B Bet on Fine-Tuning at Scale

 What’s happening: Ex-OpenAI CTO Mira Murati’s Thinking Machines just unveiled Tinker, a cloud service that industrializes fine-tuning. Instead of forcing teams to spin up full retrains, Tinker applies LoRA adapters—tiny parameter layers—so developers can spin out task-specific models without duplicating a 70B-weight beast. Multiple teams can share the same backbone, error recovery is automated, and a “Cookbook” of pre-built workflows cuts setup from weeks to hours.

 How this hits reality: If this scales, it flips the power dynamic. OpenAI and Anthropic still monetize access to frontier models, but Tinker makes open-source stacks commercially viable for anyone with a credit card. Fine-tuning at scale means startups, labs, even enterprises can deploy “good-enough” custom models without paying per-token taxes to closed platforms. Giants that rely on API lock-in suddenly look exposed.

 Key takeaway: If Tinker works, the moat isn’t model size, it’s who commoditizes customization.

Tinker is a flexible API for efficiently fine-tuning open source models with LoRA. It's designed for researchers and developers who want flexibility and full control of their data and algorithms without worrying about infrastructure management. 

LoRA is an efficient approach to fine-tuning that trains a streamlined adapter instead of updating all base model weights. Our research demonstrates that with the right setup, LoRA matches the learning performance of full fine-tuning while providing more flexibility and requiring less compute. 

https://thinkingmachines.ai/tinker/

https://thinkingmachines.ai/blog/defeating-nondeterminism-in-llm-inference/

25 de setembro de 2025, EB

https://medium.com/neo4j/why-im-replacing-my-rag-based-chatbot-with-agents-1718ec7e384c

“The rumors of RAG’s demise have been greatly exaggerated."

That was the response from Alexy Khrabrov, Neo4j’s eyes and ears on the ground in San Francisco, when I asked him whether I should be worried about our GenAI curriculum on GraphAcademy.

That said, public opinion has turned on retrieval-augmented generation (RAG), and I can’t move for thought pieces on how RAG is dead. I see this as mostly hyperbole, but I do think the position of RAG within a GenAI stack is changing.

24 de setembro de 2025, EB

https://medium.com/coding-nexus/googles-embeddinggemma-a-lightweight-open-source-embedding-model-you-can-run-anywhere-b218d4b4ac9d

Google’s EmbeddingGemma: A Lightweight, Open-Source Embedding Model You Can Run Anywhere

EmbeddingGemma takes a different path. It’s a small, open-source embedding model (308M parameters) that runs right on your laptop, desktop, or even your phone. No servers. No internet. Just you and the model.

https://ai.google.dev/gemma/docs/embeddinggemma/inference-embeddinggemma-with-sentence-transformers#:~:text=EmbeddingGemma%20is%20a%20lightweight%2C%20open,with%20no%20internet%20connection%20required.

https://developers.googleblog.com/ja/introducing-embeddinggemma/#:~:text=Gemma-,Introducing%20EmbeddingGemma:%20The%20Best%2Din%2DClass%20Open,Model%20for%20On%2DDevice%20Embeddings&text=We're%20excited%20to%20introduce,models%20nearly%20twice%20its%20size.

EmbeddingGemma, a new open embedding model that delivers best-in-class performance for its size. Designed specifically for on-device AI, its highly efficient 308 million parameter design enables you to build applications using techniques such as Retrieval Augmented Generation (RAG) and semantic search that run directly on your hardware. It delivers private, high-quality embeddings that work anywhere, even without an internet connection. 

Built to flexibly work offline: Small, fast, and efficient, it offers customizable output dimensions (from 768 to 128 via Matryoshka representation) and a 2K token context window to run on everyday devices like mobile phones, laptops, desktops, and more. Designed to work with Gemma 3n, together they unlock new use cases for mobile RAG pipelines, semantic search, and more.

Integrated with popular tools: To make it easy to get started with EmbeddingGemma, it already works with your favorite tools, such as sentence-transformers, llama.cpp, MLX, Ollama, LiteRT, transformers.js, LMStudio, Weaviate, Cloudflare, LlamaIndex, LangChain and more! 

12 de setembro de 2025, EB

https://medium.com/data-science-in-your-pocket/google-gemma3-270m-the-best-smallest-llm-for-everything-efcf927a74be

Google Gemma3 270M : The best Smallest LLM for everything

How to use Google Gemma3 270M for free?

This isn’t just a lightweight model, it’s a tool built to do real work, efficiently, without draining your device or your wallet.

But don’t let the number fool you, Gemma 3 270M isn’t built to win size contests. It’s built to follow instructions, run on tiny devices, and get fine-tuned for laser-focused tasks.

Gemma 3 270M works on this same logic, it’s perfect when you need one thing done really well. Things like:

Turning messy text into structured data

Classifying emails or support tickets

Extracting entities from legal documents

Filtering toxic content in a multilingual app

Powering simple creative tools like a bedtime story generator

Why waste compute running a multi-billion-parameter model on that?

https://medium.com/coding-nexus/googles-new-llm-runs-on-just-0-5-gb-ram-here-s-how-to-fine-tune-it-locally-ab910fa39732

In a world chasing bloat, Gemma 3 270M is a return to minimalism, with brains. If you’ve got a focused task, don’t over-engineer it. Start small. Start with Gemma 3 270M.

https://blog.stackademic.com/i-built-a-fully-offline-ai-agent-that-answers-questions-from-pdf-images-and-audio-no-cloud-2e8d71b246d6

I trained my own local RAG system with OCR, Whisper, CLIP, and open-source language models — all running on a mid-tier laptop. 

22 de agosto de 2025, EB

https://blog.gdeltproject.org/building-a-story-index-of-a-decade-of-tv-evening-news-with-geocoding-sentiment-frame-narrative-analysis-in-one-hour-for-154/

Building A Story Index Of A Decade Of TV Evening News With Geocoding, Sentiment, Frame & Narrative Analysis In 1 Hour For $154

https://blog.gdeltproject.org/

https://www.gdeltproject.org/

30 de julho de 2025, EB

https://mail.google.com/mail/u/0/#inbox/FMfcgzQbgcKklJwGzLkPfRnTXCbGwrxH

But by answering more questions directly—whether through featured snippets, knowledge panels, or now AI summaries—Google transitions from being a mere index to an oracle.

Advertisers, desperate for access to the dwindling pool of users who actually leave Google’s domain, find themselves bidding up the price. For Google, this would be a masterstroke—extracting greater rents from advertisers while providing users with what appears to be a more efficient search and knowledge acquisition experience.

Their post-click behavior becomes a crucial determinant of value for advertisers

This dynamic has analogues in the past history of digital advertising: consider the shift from indiscriminate banner ads to targeted, intent-driven search ads in the early 2000s, which dramatically increased both the value-per-click and the efficacy of online marketing.

Google’s response: if anyone is going to disrupt our business by doing this, it is going to be us.

We will thus cannibalize its own legacy business and simultaneously define the next paradigm—one in which the search page itself becomes the user’s destination. One thinks of IBM’s embrace of the PC, or Apple’s willingness to let the iPhone kill the iPod

Imagine a world in which one confronts not a list of blue links, but a cogent synthesis, drawing on JSTOR, Wikipedia, and the latest Substack polemic, all in one go. Frictionless knowledge, delivered at the speed of thought, with the AI acting as both librarian and tutor.

But the losers in this brave new world are the websites.

 The web risks becoming a substrate for AI extraction, rather than a living network of destinations. And then we get a new form of SEO, dedicated to maximizing AI-slop.

But there is a price: obscured nuance, flattened debate, and consensus prioritized over dissent. Moreover, there is the loss of “productive friction”—the serendipity and discovery that comes from wandering through primary sources.

Of course, this strategy risks eroding the web’s ecosystem of content creators, fragmenting user journeys, and triggering regulatory backlash.

the long-term consequences for information quality, diversity, and innovation remain uncertain.

Could publishers respond by designing landing pages specifically for “AI-primed” visitors—pages that acknowledge the summary, offer deeper context, or present unique value not captured in the initial answer

Might we see a bifurcation of the web: one layer optimized for human navigation and engagement, another for AI ingestion and synthesis?

How can users maintain agency, critical thinking, and serendipitous discovery in a world where the MAMLM’s output is tailored, persuasive, and frictionless?

10 de julho de 2025, EB

https://arxiv.org/html/2407.16833v1

 While LC demonstrate superior performance in long-context understanding, RAG remains a viable option due to its lower cost and advantages when the input considerably exceeds the model’s context window size. Our proposed method, which dynamically routes queries based on model self-reflection, effectively combines the strengths of both RAG and LC, achieving comparable performance to LC at a significantly reduced cost. We believe our findings contribute valuable insights for the practical application of long-context LLMs and pave the way for future research in optimizing RAG techniques.

https://www.jellyfishtechnologies.com/building-rag-agent-with-google-adk-and-vertex-ai/

23 de junho de 2025, EB

https://machine-learning-made-simple.medium.com/fine-tuning-llms-is-a-huge-waste-of-time-bd0b98fcc282

Fine-tuning isn’t knowledge injection — it’s knowledge overwrite. For advanced LLMs, neurons are no longer neutral placeholders; they’re highly specialized, densely interconnected repositories of valuable information. Carelessly updating them risks catastrophic, invisible damage.
If your goal is to build adaptable, scalable, and robust systems, treat fine-tuning with the caution it deserves. Embrace modular solutions (software principles don’t dissapear just b/c we’re working on AI) that maintain the integrity of your network’s foundational knowledge. Otherwise, you’re simply dismantling your carefully constructed knowledge ecosystem — one neuron at a time.

12 de junho de 2025, EB

Aprender com os prompts "leaked" dos outros:

https://simonwillison.net/2025/May/25/claude-4-system-prompt/
https://github.com/asgeirtj/system_prompts_leaks/blob/main/OpenAI/app/o3.md
https://github.com/jujumilk3/leaked-system-prompts

2 de junho de 2025, EB

Pinpoint Updates: Faster audio transcription and new AI features
You can now transcribe audio up to 20 times faster! We've also added automatic language detection, so you can breeze through your audio files without needing to specify the language beforehand. This enhancement will significantly speed up your workflow.
The team is also introducing time-saving generative AI features, available to early access users:

• Effortless Labeling: You can now automatically label up to 100 documents at a time by simply setting your criteria for Pinpoint to follow.

• Streamlined Data Extraction: Need specific information from multiple documents? You can now extract data from multiple documents by telling Pinpoint what you need. The results will be delivered in an easy-to-use spreadsheet, complete with links back to the original text.

• Quick Collection Summaries: Get the gist of your entire collection or specific documents in a flash with our new summarization feature.

• Easy Document Comparison: Quickly identify key differences and similarities by comparing up to three documents within a collection side-by-side.

https://mail.google.com/mail/u/0/#inbox/FMfcgzQbfVBKJrfwWGTBLlNtCqMBsbbb 

2 de junho de 2025, EB

Microsoft’s hybrid AI architecture represents a monumental shift in PC computing, fostering an innovative ecosystem that encourages the development of AI-driven applications. While challenges remain, the potential for a more intelligent and adaptive operating system is significant 

Microsoft’s new Windows architecture dynamically distributes AI tasks between local neural processing units (NPUs) and the cloud, optimizing performance, enhancing user privacy, and improving energy efficiency.

The Windows AI Foundry serves as an ecosystem for developers to create, manage, and deploy AI models. It features over 1,900 optimized AI models, end-to-end workflow tools, and strong security compliance measures, making AI development more accessible and efficient.

Microsoft envisions Windows as a platform for autonomous agents capable of performing complex tasks independently, enhancing user productivity while safeguarding data privacy. The new standards facilitate communication between these agents, enabling them to collaborate in achieving user goals.

https://aisecret.us/duolingos-ai-torn-apart/?ref=ai-secret-newsletter 

4 de abril de 2025, EB

Biologia de um LLM:
we don’t understand how models do most of the things they do.
Knowing how models like Claude think would allow us to have a better understanding of their abilities, as well as help us ensure that they’re doing what we intend them to. For example:

• Claude can speak dozens of languages. What language, if any, is it using "in its head"?

• Claude writes text one word at a time. Is it only focusing on predicting the next word or does it ever plan ahead?

• Claude can write out its reasoning step-by-step. Does this explanation represent the actual steps it took to get to an answer, or is it sometimes fabricating a plausible argument for a foregone conclusion?
  
  

https://www.anthropic.com/research/tracing-thoughts-language-model?utm_source=substack&utm_medium=email

https://transformer-circuits.pub/2025/attribution-graphs/biology.html

https://mail.google.com/mail/u/0/#inbox/FMfcgzQZTzXtxNHzCzDljKhxmdwSCtdJ?compose=CllgCJTNqGGftdXqNtbnqfCJXvCSnmwLpnmSKNpSbgbkTcszwwDvxPZWlzrjbxLfhNjWhNtTWWg 

11 de março de 2025, JCA

Este tweet é interessante:

https://x.com/WesternMishima/status/1897726036470907315 

Mistral OCR não foi bom para manuscritos (ex: escrita do Salazar), mas GPT 4.5 (lançado há umas semanas) funcionou bem.

28 de fevereiro de 2025, EB

https://research.ibm.com/blog/docling-generative-AI

Docling, IBM’s new open-source toolkit, is designed to more easily unearth that information for generative AI applications. The toolkit streamlines the process of turning unstructured documents into JSON and Markdown files that are easy for large language models (LLMs) and other foundation models to digest. Once machine-readable, this data can be used to train and customize AI models and AI agents for enterprise tasks and to ground them on the latest, most accurate facts through retrieval-augmented generation (RAG).

Docling features a command-line interface, a Python API, and is small enough to run on a standard laptop. It takes just five lines of code to set up, and it integrates seamlessly with open-source LLM frameworks like LlamaIndex and LangChain for RAG and question-answering applications. Its permissive, open-source MIT license allows developers to collaborate and expand the project to meet their needs. Docling sidesteps OCR when it can, in favor of computer vision models trained to recognize and categorize the visual elements on a page.

16 de fevereiro de 2025, JCA

Um benchmark e análise do novo modelo da Google, Gemini Flash 2, para extrair conteúdo de PDFs. $1 USD por ~6.000 páginas e 0.84 de precisão, é um preço e resultado excelente.

https://www.sergey.fyi/articles/gemini-flash-2 

15 de janeiro de 2025, EB

In contrast, China follows a state-driven, diverse AI development plan. Like the United States, China also invests in LLMs but simultaneously pursues alternate paths to GAI, including those more explicitly brain-inspired. This report draws on public statements by China’s top scientists, their associated research, and on PRC government announcements to document China’s multifaceted approach.

The Chinese government also sponsors research to infuse “values” into AI intended to guide autonomous learning, provide AI safety, and ensure that China’s advanced AI reflects the needs of the people and the state. This report concludes by recommending U.S. government support for alternative general AI programs and for closer scrutiny of China’s AI research.

New approaches such as training not based on human feedback but using different value models might be suitable for specific, well-defined scenarios, but it is unclear how such a strategy could generalize to broader sets of ethical dilemmas, let alone lead to LLM responses that are consistent with a specific set of values. It is therefore currently far from clear if and how particular sets of values can be “trained into” an LLM, given that, for LLMs, “good” and “bad” are just words to be predicted, with no grounding in any kind of valuative framework.

A final distinction between Chinese and western AI research evidenced throughout this paper needs to be made explicit, namely, the real possibility that China’s directed, strategic approach may be more effective—all other things being equal—than the western profit-driven approach that focuses on quick wins at the possible expense of more successful strategies that require a longer time horizon

https://cset.georgetown.edu/publication/chinese-critiques-of-large-language-models/?utm_source=Center+for+Security+and+Emerging+Technology&utm_campaign=3f3e8aec6f-Chinese+Critiques+of+Large+Language+Models&utm_medium=email&utm_term=0_fcbacf8c3e-3f3e8aec6f-438422560 

3 de setembro de 2024, EB

I never liked writing with Grammarly correcting me in the moment because that interferes with my flow. I didn’t like pasting my text in after finishing because it annoyed me to look through the text for errors, and it especially annoyed me when Grammarly wanted to change something for no good reason at all.
When ChatGPT arrived, I switched to using it and have been much happier with the results. Here are the instructions I gave ChatGPT before pasting in my text.

https://medium.com/tech-and-me/grammarly-vs-chatgpt-for-proofreading-ac1d7e5008f3 

14 de maio de 2024, EB

From The New York Times:
OpenAI Unveils New ChatGPT That Listens, Looks and Talks
Chatbots, image generators and voice assistants are gradually merging into a single technology with a conversational voice.

https://www.nytimes.com/2024/05/13/technology/openai-chatgpt-app.html?smid=em-share

12 de março de 2024, EB

From The New York Times:

Now, a start-up founded by three former OpenAI researchers is using the technology development methods behind chatbots to build A.I. technology that can navigate the physical world.

Covariant, a robotics company headquartered in Emeryville, Calif., is creating ways for robots to pick up, move and sort items as they are shuttled through warehouses and distribution centers. Its goal is to help robots gain an understanding of what is going on around them and decide what they should do next.

The technology also gives robots a broad understanding of the English language, letting people chat with them as if they were chatting with ChatGPT.

The technology, still under development, is not perfect. But it is a clear sign that the artificial intelligence systems that drive online chatbots and image generators will also power machines in warehouses, on roadways and in homes.

https://www.nytimes.com/2024/03/11/technology/ai-robots-technology.html 

14 de dezembro de 2023, EB

ChatGPT to summarize Politico and Business Insider articles in ‘first of its kind’ deal

OpenAI to pay German media group Axel Springer to use its material, including stories behind paywalls

https://www.theguardian.com/technology/2023/dec/13/openai-axel-springer-chatgpt-story-writing-business-insider-politico 

19 de outubro de 2023, EB

Digital artist Daniel Ambrosi has created an exhibition that interprets quintessentially English, eighteenth-century vistas with AI

https://www.theguardian.com/technology/gallery/2023/oct/09/ai-and-the-landscapes-of-capability-brown-in-pictures?CMP=share_btn_link 

12 de setembro de 2023, EB

From FT:
    Generative AI exists because of the transformer. This is how it: Writes, Works, Learns, Thinks and Hallucinates

https://ig.ft.com/generative-ai/

8 de Agosto de 2023, EB
AI can identify passwords by sound of keys being pressed, study suggests

Researchers create system using sound recordings that can work out what is being typed with more than 90% accuracy

https://www.theguardian.com/technology/2023/aug/08/ai-could-identify-passwords-by-sound-of-keys-being-pressed-study-suggests?CMP=share_btn_link 

12 de junho de 2023, JCA

Novidades de Machine Learning da Apple na WWDC, incluido o seu software com UI, que não necessita de código:

• https://developer.apple.com/machine-learning/create-ml/ 

Versão 2 de um produto de geração de vídeo via prompts, dos mesmos investigadores que criaram o Stable Diffusion:

• https://research.runwayml.com/gen2

15 de maio de 2023, JCA

Um bom guia, extraído da experiência interna de uma empresa, sobre Large Language Models e conselhos sobre ‘prompts’:

• https://github.com/brexhq/prompt-engineering 

9 de maio de 2023, AA

Sobre coleções de antropologia colonial

• https://www.publico.pt/2023/05/07/culturaipsilon/noticia/esqueleto-humano-nao-objecto-museu-2048630 

21 de abril de 2023, EB

• Como funcionam os LLM e o GPT: magnifico (mas longo) ensaio de Stephen Wolfram: https://writings.stephenwolfram.com/2023/02/what-is-chatgpt-doing-and-why-does-it-work (pdf). 

• Ver também no The Economist: "Large, creative AI models will transform lives and labour markets": https://www.economist.com/interactive/science-and-technology/2023/04/22/large-creative-ai-models-will-transform-how-we-live-and-work (pdf)

• Ligação OpenGPT com Wolfram: "ChatGPT + Wolfram is something very new—really a completely new kind of technology.": https://writings.stephenwolfram.com/2023/03/chatgpt-gets-its-wolfram-superpowers (pdf)

• Sobre implicações dos LLM e apps generativas, ver o excelente ensaio no The Economist: "How AI could change computing, culture and the course of history", https://www.economist.com/essay/2023/04/20/how-ai-could-change-computing-culture-and-the-course-of-history  (pdf) (keep)

29 de Março de 2023, EB

• • "What has actually been achieved on this video call? It takes Jared Spataro just a few clicks to find out. Microsoft’s head of productivity software pulls up a sidebar in Teams, a video-conferencing service. There is a 30-second pause while somewhere in one of the firm’s vast data centres an artificial-intelligence (ai) model analyses a recording of the virtual meeting so far. Then an impressively accurate summary of your correspondent’s questions and Mr Spataro’s answers appears. Mr Spataro can barely contain his enthusiasm. “This is not your daddy’s ai,” he beams."

  • https://www.economist.com/business/2023/03/26/big-tech-and-the-pursuit-of-ai-dominance?itm_source=parsely-api

  • Potencial do GPT para resumir entrevistas e reuniões.

  • Surpreendente potencial multilingua do GPT

3 de Março de 2023, JCA

• • Novas API para OpenGPT e Whisper (https://openai.com/blog/introducing-chatgpt-and-whisper-apis)

  • Customização do modelo GPT para datasets especificos (https://openai.com/blog/customizing-gpt-3) . 

  • Ver ELICIT, The AI research assistant (https://elicit.org) 

27 de Janeiro de 2023, AA

Tecnologias OCR. digitalização de jornais

• • https://ocr-d.de/en  (OCR Development, DFG project)

  • https://www.newseye.eu  (NewsEye. A Digital Investigator for Historical Newspapers. EU project)

21 de setembro de 2022, JCA

"Speech recognition", identificação de palavras em ficheiros de audio, exemplos de aplicação Whisper da OpenAI (https://openai.com/research/whisper )

• https://box.wolan.net/plenario-portugal-old.mp4 

• https://box.wolan.net/assembleia_municipal_taxa.mp4 

30 de julho de 2022, EB

Reconstrução histórica das notícias sobre “Carrazeda de Ansiães” (um concelho do vale do Tua, o “interior” do interior do norte de Portugal) a partir das ocorrências nos dois periódicos entre 1883 e 1893, parte de uma investigação em curso sobre a história da região nos finais do século XIX (a ortografia dos textos originais foi corrigida com a ajuda do corretor do Google Docs)

• https://docs.google.com/document/d/1cn45SWWlXYtmWTzO3NUOuXDnb7a8GAEyVAs15mGskDA/edit?usp=sharing 

20 de julho de 2022, AA

Notas sobre o layout das fichas da Botânica (mhnc):

• https://docs.google.com/document/d/1APlXXlII7Y0GdJTQYMgTZvbjYgMw-5mzewjgEh0ZrxA/edit?usp=sharing

Tabelas com as características das imagens das 3 gavetas digitalizadas da Botânica (mhnc):

• https://docs.google.com/spreadsheets/d/1CGDhFjJW2kmkgxrHg51OBOZLCsk8qfzr6zegazJa6y4/edit?usp=sharing

12 de julho de 2022, AA

Notas sobre o layout das fichas da Botânica (mhnc):

• https://docs.google.com/document/d/1APlXXlII7Y0GdJTQYMgTZvbjYgMw-5mzewjgEh0ZrxA/edit?usp=sharing

Tabelas com as características das imagens das 3 gavetas digitalizadas da Botânica (mhnc):

• https://docs.google.com/spreadsheets/d/1CGDhFjJW2kmkgxrHg51OBOZLCsk8qfzr6zegazJa6y4/edit?usp=sharing

5 de julho de 2022, AA

Revisão do estado da arte de 2022, dos vários serviços cloud, para extração de informação de fichas:

• https://docs.google.com/document/d/1VKDVBsi9v4TLHGSPqLE_KX0DvW5mqZl0CP9PGiZ8nrc/edit?usp=sharing 

5 de julho de 2022, EB

Avaliação da correcção de erros pelo corrector ortográfico do google docs:

• • https://docs.google.com/document/d/1Ewg1_aIKwXWGAkZoWlIguqjsW-B3wkSURUF3nKSoSZA/edit?usp=sharing

5 de julho de 2022, AA

Análise  de performance dos vários serviços de ocr nas fichas do Ateneu:

• • https://docs.google.com/document/d/1h_2PRf6UpgW6RGBOEOJTRls3YKbFKgEt4yFsgni1Cf0/edit?usp=sharing

  • https://docs.google.com/document/d/1VKDVBsi9v4TLHGSPqLE_KX0DvW5mqZl0CP9PGiZ8nrc/edit?usp=sharing

  • https://docs.google.com/spreadsheets/d/1R5w3nAGFfNIgoGkWGzhmLMCfFr89ELYTPoe9bPVhjfk/edit?usp=sharing

20 de junho de 2022, JCA

A framework for designing document processing solutions

• • https://docs.google.com/document/d/1nttjhsdPdkJZWfIXx1Oxd39883tgG7PBcp9VC1ayF6w/edit?usp=sharing 

11 de fevereiro de 2022, EB

Tutorial: Beginner’s guide to machine learning in R

• • https://docs.google.com/document/d/1nttjhsdPdkJZWfIXx1Oxd39883tgG7PBcp9VC1ayF6w/edit?usp=sharing 

15 de outubro de 2021, AA

diagramas de fluxo para os processos de geração de txt, ocr e hocr dos jornais

• • https://drive.google.com/file/d/1Y7pLPel0pueeqo2konBtuBKOvJEIMrGF/view?usp=share_link

  • https://drive.google.com/file/d/1w2eA_Xgfdua69UI7gERhLr3W9aSvoERc/view?usp=share_link 

29 de julho de 2021, AA

papers de estratégias de correção de erros de jornais finlandesas

• • https://drive.google.com/file/d/1dP_ZiM8ae716V2E_8wZO_Y5OHkv1hnZ2/view 

  • https://drive.google.com/file/d/111UTI6ELaU2Wca-Yn_IpJoECGtFx6Ozs/view 

19 de abril de 2021, JCA

hOCR tools

• •  https://github.com/ocropus/hocr-tools