This week we feature articles by Amber Case, and Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Yarin Gal, Nicolas Papernot, & Ross Anderson.

Additional reading comes from Raiza Martin & Steven Johnson, Jerry Liu, Sara Fischer, and Nikhil Simha.

News comes from Adobe, CrafterCMS, and Bloomreach.

👉 It’s time for our annual summer break. Our next issue will be published on September 6. I hope you all have a pleasant, and not-too-hot-or-rainy, August.

All previous issues are available at https://gilbane.com/gilbane-advisor-index

---

Opinion / Analysis

The curse of recursion: Training on generated data makes models forget

Training on generated data is an iffy proposition at best, yet it is also difficult to avoid, or even detect. This is not an unknown problem, but is certainly under-appreciated. This paper argues that the problem is unavoidable and explains why. There is lots of math but the narrative is well written, making the paper useful even for the math-squeamish. (28 min)

“Model Collapse is a degenerative process affecting generations of learned generative models, where generated data end up polluting the training set of the next generation of models; being trained on polluted data, they then mis-perceive reality.”…

Abstract: https://arxiv.org/abs/2305.17493v2 Full PDF: https://arxiv.org/pdf/2305.17493v2.pdf

How to identify “Truthy” tech trends

Amber Case provides an entertaining, and at least mostly true, analysis of truthiness in tech. It’s also likely that most of us know multiple people who would benefit from reading it. (8 min)

https://caseorganic.medium.com/how-to-identify-truthy-tech-trends-70f553c5a445

More Reading

• “An AI-first notebook, grounded in your own documents” Introducing NotebookLM via Google Labs
• LLM-powered knowledge workers… Data Agents via LlamaIndex blog
• AP strikes news-sharing and tech deal with OpenAI via Axios
• Case study… Chronon — A declarative feature engineering framework via The Airbnb Tech Blog

All Gilbane Advisor issues

---

Content technology news

Adobe Firefly supports prompts in over 100 languages

Users can now generate high-quality images, create text effects, streamline workflows and improve productivity in their language of choice.
https://firefly.adobe.com

CrafterCMS release version 4.1

The updated CMS includes several new content authoring features for creating content-centric digital experiences, and a switch to OpenSearch.
https://craftercms.com

Bloomreach supports OAuth 2.0 authentication

Businesses have the ability to integrate Bloomreach Engagement with third-party applications that require OAuth through webhooks.
https://www.bloomreach.com/en/news/2023/bloomreach-now-supports-frictionless-integrations-with-third-party-applications-requiring-oauth-2.0-authentication-method

All content technology news

---

The Gilbane Advisor is authored by Frank Gilbane and is ad-free, cost-free, and curated for content, computing, web, data, and digital experience technology and information professionals. We publish recommended articles and content technology news weekly. We do not sell or share personal data.

Subscribe | View online | Editorial policy | Privacy policy | Contact

This week we feature an article by Boaz Barak.

Additional reading comes from Anna Kaley, Maria Rosala, Kara Pernice & Patty Caya, Rachel Andrew, Tom Davenport & Maryam Alavi, and Dean Allemang.

News comes from Brightspot, dtSearch, Databricks, and Snowflake.

All previous issues are available at https://gilbane.com/gilbane-advisor-index

---

Opinion / Analysis

Metaphors for AI, and why I don’t like them

Boaz Barak finds metaphors very useful tools for facilitating understanding, but as Dennett ‘intuition pumps’ that leave out complicating detail. Barak is concerned that the use of metaphors to explain AI is leading to “over-interpretation”. In this excellent piece he looks at the ‘stochastic parrot’, ‘blurry JPEG of the web’, and other AI metaphors to make his case for caution. (13 min)

https://windowsontheory.org/2023/06/28/metaphors-for-ai-and-why-i-dont-like-them/

More Reading

• 10 best intranets of 2023: What makes them great via Nielsen Norman Group
• New to the web platform in May and June via web .dev
• Good read for senior managers… How to train Generative AI using your company’s data via HBR .org
• Sometimes it looks like they do… LLM, explain yourself! via Dean Allemang

All Gilbane Advisor issues

---

Content technology news

dtSearch updates enterprise products

dtSearch enterprise and developer product line searches terabytes across online and offline data; products can run on premises or the cloud.
https://www.dtsearch.com

Brightspot integrates OpenAI

Content creators can leverage AI-assisted suggestions for headlines, subheadlines and full body text to expedite the content-creation process.
https://www.brightspot.com/cms-resources/tag/artificial-intelligence

Databricks announces LakehouseIQ

LakehouseIQ is a knowledge engine using generative AI to provide useful, contextual answers to natural language questions from all approved employees.
https://www.databricks.com/blog/introducing-lakehouseiq-ai-powered-engine-uniquely-understands-your-business

Snowflake unveils large language model to extract data from documents

Document AI leverages Snowflake’s first-party large language model to unlock deeper insights from unstructured data in documents.
https://www.snowflake.com/news/snowflake-unveils-new-large-language-model-to-extract-deeper-insights-from-documents-while-continuing-to-advance-platform-speed-and-performance/

All content technology news

---

The Gilbane Advisor is authored by Frank Gilbane and is ad-free, cost-free, and curated for content, computing, web, data, and digital experience technology and information professionals. We publish recommended articles and content technology news weekly. We do not sell or share personal data.

Subscribe | View online | Editorial policy | Privacy policy | Contact

This week we feature articles from Shirui Pan, Linhao Luo, Yufei Wang, Chen Chen, Jiapu Wang & Xindong Wu, and Maria Antoniak, Li Lucy, Maarten Sap & Luca Soldaini.

Additional reading comes from Jon Udell, Fabio Matricardi, Mark Sundstrom, and Rhiannon Williams.

News comes from MongoDB, Drupal, Databricks, and Bloomreach & Shopware.

Note: We’ll be off next week and back on July 12th.

All previous issues are available at https://gilbane.com/gilbane-advisor-index

---

Opinion / Analysis

Unifying large language models and knowledge graphs: A roadmap

We’ve been focused mostly on practitioner-authored articles on this topic. This comprehensive paper is a useful resource for practitioners, strategists, and researchers alike. It is well written and illustrated, and a valuable reference that will come in handy at multiple points in related projects. (50 min)

https://arxiv.org/pdf/2306.08302.pdf

Using large language models with care

This is not news, but researchers Maria Antoniak, Li Lucy, Maarten Sap, and Luca Soldaini from The Allen Institute provide an important public service with their “introductory outline of the risks of LLMs, written for the everyday user.” This will be directly helpful to some of us, but especially useful for sharing with certain colleagues, or family and friends. (11 min)

https://blog.allenai.org/using-large-language-models-with-care-eeb17b0aed27

More Reading

• Why LLM-assisted table transformation is a big deal Some of you will know this – either way, a useful short read. via Jon Udell
•  Fun take on hype cycles for programmers… XML is the future via Bite code!
• For adventurous DIY’s like Fabio Matricardi… Answering questions about your documents using LangChain (and not OpenAI) via Artificial Corner
• Sad, but inevitable… The people paid to train AI are outsourcing their work… to AI via MIT Technology Review

All Gilbane Advisor issues

---

Content technology news

MongoDB launches five new capabilities for MongoDB Atlas

The new MongoDB features can help speed up innovation by standardizing workloads on a single developer data platform across the enterprise.
https://www.mongodb.com/press/mongodb-launches-five-new-capabilities-for-mongodb-atlas-to-build-new-classes-of-applications

Drupal 10.1 is now available

Customizing the look of your site easier, adds support for decoupled navigation, improves content modeling and editing, block management…
https://www.drupal.org/project/drupal/releases/10.1.0

Databricks announces Lakehouse Apps

Databricks customers will be able to discover, deploy, and manage data and AI applications directly from Databricks.
https://www.databricks.com

Bloomreach and Shopware partner

Bloomreach’s marketing automation, product discovery, and content management will enhance Shopware e-commerce customer’s customer journeys.
https://www.bloomreach.com ■ https://www.shopware.com/en/

All content technology news

---

The Gilbane Advisor is authored by Frank Gilbane and is ad-free, cost-free, and curated for content, computing, web, data, and digital experience technology and information professionals. We publish recommended articles and content technology news weekly. We do not sell or share personal data.

Subscribe | View online | Editorial policy | Privacy policy | Contact

This week we feature articles from Simon Wardley, and Jakob Nielsen.

Additional reading comes from Sayash Kapoor & Arvind Narayanan, Rishi Bommasani, Kevin Klyman, Daniel Zhang & Percy Liang, and Alex Ratner.

News comes from Acquia, Adobe, Optimizely, and Expert[.]ai & Blue Prism.

All previous issues are available at https://gilbane.com/gilbane-advisor-index

---

Opinion / Analysis

Both articles this week argue that the current state of conversational programming with prompt engineering as the interface, is primitive, though directional. Simon Wardley’s post is a bit more technical, while Jakob Nielsen’s focuses on end-user usability.

Why the fuss about conversational programming

“In my last post on conversational programming I asked the reader to “get yourself ready for a world of conversational programming” but I wasn’t willing to call it for ChatGPT or the existing crop of LLMs.”

If you’re familiar with Simon Wardley, you won’t be surprised that he says we need to get to “code as maps”. Otherwise, see his previous post and perhaps some of his other links. (10 min)

https://swardley.medium.com/why-the-fuss-about-conversational-programming-70a8b7ca0d2b

AI: first new UI paradigm in 60 years

Jakob Nielsen…

“I doubt that the current set of generative AI tools (like ChatGPT, Bard, etc.) are representative of the UIs we’ll be using in a few years, because they have deep-rooted usability problems. Their problems led to the development of a new role — the prompt engineer.” (5 min)

https://www.nngroup.com/articles/ai-paradigm/

More Reading

• It’s not so simple… Is AI-generated disinformation a threat to democracy? via AI Snake Oil
• Handy comparison table with details and scores for each requirement… Do foundation model providers comply with the draft EU AI Act? via Stanford Center for Research on Foundation Models
• Of course we need a platform… Snorkel’s Foundation Model Data Platform via Snorkel

All Gilbane Advisor issues

---

Content technology news

Acquia enhances digital experience platform for personalization

Enables digital experiences based on knowledge gained from a broader array of channels, using anonymous, known, and understood customer data.
https://www.acquia.com/newsroom/press-releases/acquia-enhances-digital-experience-platform-power-omnichannel-customer

Expert[.]ai expands partnership with SS&C Blue Prism

Processing unstructured language data for automation creates new potential for leveraging intelligent automation in the enterprise.
https://www.expert.ai ■ https://www.blueprism.com

Optimizely introduces Content Graph

Content Graph service extends the capabilities of Optimizely’s CMS by enabling flexible content delivery across channels and applications.
https://www.optimizely.com/headless

Adobe adds Firefly generative AI capabilities to Illustrator

Following Adobe Photoshop and Adobe Express, Illustrator is the latest Creative Cloud application to benefit from Adobe’s generative AI.
https://www.adobe.com/products/illustrator.html

All content technology news

---

The Gilbane Advisor is authored by Frank Gilbane and is ad-free, cost-free, and curated for content, computing, web, data, and digital experience technology and information professionals. We publish recommended articles and content technology news weekly. We do not sell or share personal data.

Subscribe | View online | Editorial policy | Privacy policy | Contact

This week we feature articles from Jacob Steinhardt, and Fan Li.

Additional reading comes from Shirley Lu & George Serafeim, Cassie Kozyrkov, M.G. Siegler, and Jeffrey Pound et al.

News comes from Adobe, Neo4j, Bloomreach, and Optimizely.

All previous issues are available at https://gilbane.com/gilbane-advisor-index

---

Opinion / Analysis

What will GPT-2030 look like?

Of course we would all like to know, but given the speed of new developments in machine learning it’s difficult to make, or have confidence in, any predictions. Jacob Steinhardt isn’t afraid to try, and his post is a must-read for anyone looking for some data points to inform their own research and thinking. You’ll want a cup of coffee or tea and perhaps a little more time than the specified read time (27 min).

https://bounded-regret.ghost.io/what-will-gpt-2030-look-like/

Explore OntoGPT for schema-based knowledge extraction

“The OntoGPT framework and SPIRES tool provide a principled approach to extract knowledge from unstructured text for integration into Knowledge Graphs (KGs), using Large Language Models such as GPT. This methodology enables handling complex relationships, ensures logical consistency, and aligns with predefined ontologies for better KG integration.”

Fan Li describes another open available tool that he believes overcome some limitations to the popular LangChain and LlamaIndex frameworks. The more tools the better. (7 min).

https://apex974.com/articles/ontogpt-for-schema-based-knowledge-extraction

More Reading

• How AI will accelerate the circular economy via HBR
• Linking un-structured Web documents to entities in knowledge graphs… Growing and serving large open-domain knowledge graphs via arXiv
• Platform play redux… Apple’s Vision Pro & Meta’s Vision Oh No via 500ish
• Generative AI vs enterprise AI… Why multi-purpose systems are hard to design safely via Cassie Kozyrkov

All Gilbane Advisor issues

---

Content technology news

Adobe announces Firefly for Enterprise

Adobe plans to enable businesses to custom train Firefly with their own branded assets, embedding Firefly into their own ecosystem.
https://news.adobe.com/news/news-details/2023/Adobe-Brings-Firefly-and-Express-to-Enterprises/default.aspx

Neo4j announces new integrations with generative AI features in Vertex AI

Enterprise customers can now leverage knowledge graphs with Google’s large language models to improve generative AI outcomes.
https://neo4j.com

Bloomreach enhances customer experience with Gorgias integration

Gorgias provides data about the customer experience and gives Engagement users a more in-depth understanding of the customer journey.
https://www.bloomreach.com/

Optimizely expands partnership with Contentsquare

Contentsquare enables Optimizely users to leverage heatmaps and data analyses to uncover customer insights and enhance experiments.
https://www.optimizely.com/insights/blog/optimizely-x-contentsquare-5-ways-were-better-together/

All content technology news

---

The Gilbane Advisor is authored by Frank Gilbane and is ad-free, cost-free, and curated for content, computing, web, data, and digital experience technology and information professionals. We publish recommended articles and content technology news weekly. We do not sell or share personal data.

Subscribe | View online | Editorial policy | Privacy policy | Contact

This week we feature articles from Jerry Liu, and Scott Brinker.

Additional reading comes from Sarah Perez, Rada Mihalcea, Oana Ignat & Zhijing Jin et al, Ben Thompson, and Airbyte.

News comes from PingCAP, SearchStax, Datometry & Databricks, and Snowflake & Neeva.

All previous issues are available at https://gilbane.com/gilbane-advisor-index/

---

Opinion / Analysis

The key point of Google’s “we have no moat” memo on generative AI is that ecosystems are the moat

Scott Brinker has a good read inspired by the Google “no moat” memo. He thinks the memo is “right on the money” and makes a convincing case why. (8 min)

https://chiefmartec.com/2023/06/the-key-point-of-googles-we-have-no-moat-memo-on-generative-ai-is-that-ecosystems-are-the-moat/

Combining text-to-SQL with semantic search for retrieval augmented generation

LlamaIndex is already useful for combining proprietary internal and external data and knowledge bases with LLMs to build custom data applications. LlamaIndex creator Jerry Liu describes their latest tool.

“We have created a brand-new query engine… that can query, join, sequence, and combine both structured data from your SQL database and unstructured data from your vector database in order to synthesize the final answer.” (10 min)

https://medium.com/llamaindex-blog/combining-text-to-sql-with-semantic-search-for-retrieval-augmented-generation-c60af30ec3b

More Reading

• Who’s using what, etc… State of data 2023 via Airbyte
• What to work on beyond LLMs… A PhD student’s perspective on research in NLP in the era of very large language models, via  Language and information technologies (LIT) University of Michigan
• Ben Thompson is a fan… Apple Vision via Stratechery
• WordPress.com challenges Substack with launch of paid newsletters via TechCrunch 

All Gilbane Advisor issues

---

Content technology news

PingCAP Unveils TiDB 7.1

With simplified operations and enhanced MySQL compatibility users can accelerate the productivity of developers and infrastructure engineers.
https://www.pingcap.com

SearchStax launches SearchStax for Good

Developers at non-profit organizations can build web and mobile applications without the high cost of managing search infrastructure.
https://www.searchstax.com/ss-for-good

Datometry partners with Databricks

To help enterprises accelerate migrations from data warehouses to the lakehouse without having to rewrite or redefine application code.
https://datometry.com

Snowflake acquires Neeva​

Will incorporate Neeva’s search experience that leverages generative AI across the Data Cloud to benefit customers, partners and developers.
https://www.snowflake.com/blog/snowflake-acquires-neeva-to-accelerate-search-in-the-data-cloud-through-generative-ai/

All content technology news

---

The Gilbane Advisor is authored by Frank Gilbane and is ad-free, cost-free, and curated for content, computing, web, data, and digital experience technology and information professionals. We publish recommended articles and content technology news weekly. We do not sell or share personal data.

Subscribe | View online | Editorial policy | Privacy policy | Contact

This week we feature articles from Irene Solaiman, and the Meta AI Blog.

Additional reading comes from Stephen O’Grady, Max G. Levy, and Huw Roberts, Alexander Babuta, Jessica Morley, Christopher Thomas, Mariarosaria Taddeo & Luciano Floridi.

News comes from Docugami, Acquia, Adobe, and Elastic.

All previous issues are available at https://gilbane.com/gilbane-advisor-index/

---

Opinion / Analysis

Generative AI systems aren’t just open or closed source

“Conversation around generative AI tends to focus on whether its development is open or closed. It’s more responsible to envision releases along a gradient.”

Irene Solaiman proposes a framework of six levels of access to generative AI systems. Her focus is on policy and responsible and ethical release of these systems, and her framework provides a useful foundation for thinking about release strategies, and relevant policy decisions at different points along the gradient. 

(4 min summary): https://www.wired.com/story/generative-ai-systems-arent-just-open-or-closed-source/
Full paper: https://arxiv.org/abs/2302.04844v1

Introducing speech-to-text, text-to-speech, and more for 1,100+ languages

Meta AI reports on their Massively Multilingual Speech (MMS) project which combines “wav2vec 2.0 … and a new dataset that provides labeled data for over 1,100 languages and unlabeled data for nearly 4,000 languages.”. They include a short demo, info on error rates, and links to the code and the technical paper. (7 min)

https://ai.facebook.com/blog/multilingual-model-speech-recognition/

More Reading

• All the announcements from Microsoft Build 2023 Book of News via Microsoft
• Artificial intelligence regulation in the United Kingdom: a path to good governance and global leadership? via Internet Policy Review
• Chatbots don’t know what stuff isn’t via Quanta Magazine
• The latest RedMonk programming language rankings via RedMonk

All Gilbane Advisor issues

---

Content technology news

Docugami announces integration with LlamaIndex

Combine LlamaIndex with Docugami’s Document XML Knowledge Graph to build LLM applications that connect users to their own business documents.
https://www.docugami.com/blog/llamaindex

Acquia updates Acquia DAM

Enhancements to its digital asset management platform, Acquia DAM (Widen), includes an AI chatbot to assist in creative workflows and collaboration.
https://www.acquia.com/products/acquia-dam

Adobe unveils Generative Fill for Photoshop

The beta release is Adobe’s first Creative Cloud application to deeply integrate Firefly. Document Cloud, Experience Cloud and Adobe Express to follow.
https://firefly.adobe.com/

Elastic unveils the Elasticsearch Relevance Engine​

Companies can use their structured and unstructured data to build custom generative AI apps without the cost of running LLMs.
https://www.elastic.co/enterprise-search/generative-ai

All content technology news

---

The Gilbane Advisor is authored by Frank Gilbane and is ad-free, cost-free, and curated for content, computing, web, data, and digital experience technology and information professionals. We publish recommended articles and content technology news weekly. We do not sell or share personal data.

Subscribe | View online | Editorial policy | Privacy policy | Contact

This week we feature articles from Skanda Vivek, and Alexander Savard.

Additional reading comes from Dean Allemang, Neil Clarke, Sridhahr Ramaswamy & Vivek Raghunathan, and Bogdan Arsintescu.

News comes from Ontotext, Expert[.]ai, Sinequa, and TransPerfect & Contentstack.

All previous issues are available at https://gilbane.com/gilbane-advisor-index/

---

Opinion / Analysis

When should you fine-tune LLMs?

Last week we led with an article looking at combining knowledge graphs with large language models to create high-value domain specific applications. This post from Skanda Vivek looks at another decision to make when building a domain specific application: whether to use a proprietary LLM, such as ChatGPT, fine-tune an open-source LLM, or train an LLM from scratch. Vivek includes cost considerations, but also includes a link to a post with detailed hosting costs. (7 min)

https://towardsdatascience.com/when-should-you-fine-tune-llms-2dddc09a404a

That little island changes everything

A really interesting case study from Lyft on implementing Apple’s Dynamic Island and Live Activities for a better user experience. Alexander Savard takes you through the design decisions. (10 min)

https://design.lyft.com/that-little-island-changes-everything-b89b108f45b4

More Reading

• Explaining rules (and OWL models) with LLMs ,via Dean Allemang
• The cost of dealing with generative AI created submissions. It continues… via Clarkesworld Magazine
• Neeva is shutting down its search engine. Too bad – was promising. via Neeva
• What LinkedIn’s giant knowledge graph does… How LIquid connects everything via LinkedIn Engineering

All Gilbane Advisor issues

---

Content technology news

Sinequa enhances platform for scientific search and clinical trial data

Combines deep learning and large language models for natural language understanding (NLU) with ChatGPT models through Azure OpenAI Service.Tridion unites web content, structured content and headless delivery
https://www.sinequa.com/enterprise-search-for-industries/healthcare-life-science/

Expert[.]ai launches AI platform for Life Sciences

The Platform combines industry language models and AI-based natural language capabilities transforming health and scientific data into insights.
https://www.expert.ai/solutions/life-science-pharma/

TransPerfect launches GlobalLink Connect app for Contentstack’s DXP

The GlobalLink Connect app is now available on the Contentstack Marketplace, part of Contentstack’s Composable Digital Experience Platform (DXP). 
https://www.transperfect.com/about/press/transperfect-launches-globallink-connect-app-contentstacks-composable-digital

Ontotext releases Target Discovery​

Platform provides pharma & biotech companies more efficient insight discovery, faster information retrieval, and advanced visual analytics.
https://www.ontotext.com/solutions/ontotexts-target-discovery/

All content technology news

---

The Gilbane Advisor is authored by Frank Gilbane and is ad-free, cost-free, and curated for content, computing, web, data, and digital experience technology and information professionals. We publish recommended articles and content technology news weekly. We do not sell or share personal data.

Subscribe | View online | Editorial policy | Privacy policy | Contact