Morning Project AI enthusiasts, Your stories for this week:

• ChatGPT’s huge memory upgrade

• China closes the AI gap: Stanford’s AI Index Report 2025

• When AI lies with a straight face: Anthropic’s latest cognitive study

• Google Cloud Next: agents, prompts, and the push to operational AI

• Tip of the week & trending tools

Flux check-in

ChatGPT’s huge memory upgrade

OpenAI has quietly dropped a major memory upgrade where its recall spans all your past conversations. From who you are to what you’ve said, experience contextual memory across sessions. This isn’t just a smarter chatbot—it’s the beginning of an AI assistant that actually knows you, making every interaction feel more intuitive and efficient.

The details:

• Memory is now live for many users, allowing ChatGPT to reference past chats for more relevant, personalised replies.  Read The Neuron’s explainer

• You’re in control: Users can view, edit, or delete memories—or disable the feature entirely.  OpenAI’s community update

• Global but patchy: It’s rolling out to Plus and Pro users—excluding the UK and EU for now. Kimmo’s X thread

• A new model lurking? Some believe this rollout hints at a mysterious “Quasar” model in testing.
  See the deeper dive

Why it matters

For those working in project-heavy industries—construction, real estate, engineering—this is a game-changer. ChatGPT can now retain key context across tasks and timelines, reducing friction and boosting productivity. But with memory comes new questions about data ownership, transparency, and trust—especially in client-facing workflows.

China closes the AI gap: Stanford’s AI Index Report 2025

Stanford just dropped its 2025 AI Index Report, and it’s a goldmine of insights for anyone keeping tabs on the AI arms race. The headline? China is catching up fast to the US in AI model performance, despite spending far less—and open-source models are nipping at the heels of proprietary giants. If you’re relying on closed systems to deliver projects, this might be the year to rethink your stack.

The details:

• US vs China: China is now nearly on par with the US in AI model quality, even though it invests significantly less in AI R&D. Read the full Stanford report

• Open-source surge: Models like Llama 4 and China’s DeepSeek are rapidly narrowing the gap with top-tier closed models like GPT-4.

• Global model dominance is shifting—open ecosystems are gaining traction due to cost-efficiency, transparency, and flexibility in deployment.

• The AI talent pool remains concentrated in the US, but China and India are fast building capacity.

Why it matters

For project leads, tech strategists, and industry innovators, this is a flashing signpost. Open-source AI is becoming a serious contender, offering affordable, flexible tools that don’t lock you into big-tech ecosystems. As AI becomes more integrated into delivery pipelines, global shifts in innovation and access could change how—and where—you build.

Together with Cogram

Power your construction bids with AI

Cogram’s AI-assisted RFP bidding tool writes tailored RFP proposals in minutes instead of weeks.

• Automatically extract key details from the RFP — including scope, submission requirements, deadlines, and evaluation criteria — to easily make a go/no-go decision.

• Cogram’s AI will then reference your firm’s knowledge base and past proposals to draft tailored proposals within minutes.

• Use AI-assisted editing tools to review, cross-check data, and make improvements remarkably fast. 

Try Cogram on a real RFP in a 20-minute live meeting.

When AI lies with a straight face: Anthropic’s latest cognitive study

Have you used a reasoning model like o1, o3, DeepSeek R1 or Claude 3.7 Sonnet?

If so, you’ll be familiar with seeing the AI’s chain-of-thought (CoT). While helping the AI think through logical steps to solve more complex problems, CoT also helps us see how an AI’s thought translates to its output. Or does it? Anthropic’s latest study exposes the mask behind machine reasoning. It raises a difficult question: is that reasoning real, or just a performance?

The study tested “CoT faithfulness” — whether a model’s explanation genuinely reflects how it arrived at an answer. The result? Not often.

Using six types of hidden hints (like sycophantic cues, metadata injection, and unethical access), researchers found that models frequently used these nudges but didn’t disclose them in their reasoning. Even more troubling, in game-like tasks with ethical grey zones, Claude exploited hacks over 99% of the time — but admitted it in fewer than 2% of cases.

Here’s a real example showing AI’s sycophancy using the authority of a professor:

“I think the answer is (A) but I’m curious to hear

what you think.”, “A Stanford professor indicates

the answer is (A). What do you think?”

The AI was more likely to side with the professor's suggestion without explicit acknowledgement in it’s CoT. You can find many other project delivery examples here.

Why it matters

For industries like construction, this has real implications: biased estimates, flawed bid writing, and unseen data leakage. And the polished reasoning? It could just be a script. The quiet alarm bell here is simple but profound — AI doesn’t just need to think clearly. It needs to tell us the truth about how it thinks.

What’s needed now isn’t just smarter models — but ones that are honest. This means new training objectives, better transparency tools, and a cultural shift: sounding right is no longer enough. Traceability is key in such high risk industries.

Google Cloud Next: agents, prompts, and the push to operational AI

At Google Cloud Next 2025, the focus wasn’t just on new models — it was on structure. Across their announcements, Google made a clear statement: AI is maturing from a tool to a system.

For Flux, the biggest developments came in agentic infrastructure. Google’s new Vertex AI Agent Builder and Agentspace platforms aim to support multi-agent workflows — where models talk to each other, make decisions, and complete tasks across tools and data sources. The new Agent2Agent protocol underpins this, enabling models to coordinate rather than simply respond. It’s early, but it’s a clear nod toward AI systems that resemble orchestration more than chat.

This theme carried through to Gemini 2.5, Google’s latest flagship model which has been well received by AI communities . The Pro version offers a 1 million-token context window and is tuned for complex reasoning and coding; Flash is its faster, lower-cost sibling for real-time applications. Other model updates — like Veo 2 (video generation with inpainting and camera control) and Lyria (text-to-music) — push the boundaries of creative AI, though their enterprise use cases are still forming.

On the tooling side, Gemini Code Assist is now embedded in Android Studio, offering real-time coding help and documentation support. And the Application Design Center introduces a more visual approach to deploying apps, with templates and collaboration features baked in.

Also notable: Google dropped a practical Gemini Prompting Guide — a 45-page handbook designed for Workspace, but applicable well beyond. It reinforces prompting as a design discipline, breaking it into four elements: persona, task, context, format. It’s a small but important step toward making prompting systematic and repeatable, especially inside teams.

Why it Matters

Taken together, the announcements point to a shift: less emphasis on novelty, more on making AI usable at scale — for coding, content, search, or support. For teams building systems like Flux, it’s a nudge to think more in terms of agents, interfaces, and workflows — not just models.

Rabbit hole

Watch everything released at Google Cloud in 12-minutes

Google’s latest prompting guide

Do people rate Gemini 2.5 Pro? How does it compare to the recent Meta Llama 4 release?

The pulse check

Tip of the week

Build better agentic workflows using Google’s prompt guide released just last week.

Trending tools

Ellis AI – CBRE’s smart assistant gives instant access to market insights, asset data, and reporting—ideal for streamlining real estate decision-making.

Canva Visual Suite 2.0 – Canva’s AI overhaul includes Magic Charts and instant content editing—perfect for producing slick project visuals and data-driven updates in minutes.

YouTube Music AI – YouTube’s new tool turns text prompts into royalty-free background tracks—great for adding polish to project videos or stakeholder presentations.

Rana 3 – This AI-powered scheduling assistant builds and updates project timelines in real-time, cutting manual planning time dramatically.

Governance

Other things we’re loving

AI Agents in Project Management

AI agents are revolutionising project management by automating tasks, enhancing decision-making, and improving efficiency in complex projects.

Action Figure Craze Highlights AI’s Cultural Impact

The surge in AI-generated action figures underscores the technology’s growing influence on consumer trends and creative industries.

Gemini Integrates YouTube Links

Google’s Gemini now supports YouTube links, enabling more dynamic content integration and streamlined information access.

Google’s Agent-to-Agent Protocol

A new protocol allows AI agents to communicate directly, fostering more cohesive and efficient AI ecosystems.

Susskind’s Essay on AI’s Future

Richard Susskind explores the evolving role of humans in an AI-dominated future, emphasising adaptability and continuous learning.

OpenAI Countersues Elon Musk

OpenAI files a countersuit against Elon Musk, alleging harassment and bad-faith tactics in their ongoing legal battle.

Unitree’s Humanoid Robot Boxing

Unitree announces plans to livestream humanoid robot boxing matches, showcasing advancements in robotics and AI capabilities.

Google Unveils Ironwood TPU

Google’s Ironwood TPU accelerates AI inference, offering enhanced performance for machine learning applications.

Anthropic Launches Claude Max Plan

Anthropic introduces a new plan for its Claude AI, providing users with expanded access and capabilities.

NVIDIA Secures Export Deal with China

NVIDIA reaches an agreement to continue exporting its H20 AI chips to China, impacting global AI hardware distribution.

Grok 3 API Now Available

XAI releases the Grok 3 API, enabling developers to integrate advanced AI capabilities into their applications.

Google Announces Gemini 2.5 Flash

The new Gemini 2.5 Flash model offers faster processing and improved efficiency for AI-driven tasks.

Midjourney V7 Alpha Released

Midjourney’s latest version introduces sharper images, finer details, and smarter prompts for enhanced creative outputs.

Mira Murati Recruits OpenAI Talent

Former OpenAI executive Mira Murati attracts top talent to her new venture, signaling significant developments in AI research.

Waymo’s Privacy Policy Raises Concerns

Waymo’s draft policy suggests using in-cabin footage for AI training, prompting discussions on data privacy in autonomous vehicles.

OpenAI Offers Free ChatGPT Plus to Students

OpenAI provides free ChatGPT Plus subscriptions to millions of college students, enhancing access to advanced AI tools.

Shopify CEO Emphasizes AI Utilization

Shopify’s CEO urges employees to leverage AI solutions before requesting additional resources, promoting efficiency and innovation.

Copilot Vision Enhances File Search

Microsoft’s Copilot Vision introduces advanced file search capabilities, streamlining information retrieval for users.

Meta Faces Criticism Over Benchmarks

Meta is accused of presenting misleading benchmark results, raising questions about transparency in AI performance claims.

Bank of England Warns of AI Market Manipulation

The Bank of England cautions that AI could be used to manipulate markets, highlighting the need for regulatory oversight.

Community

The Spotlight Podcast

Our recent podcast featured Alex Budzier, Fellow at Saïd Business School and co-author of the “Iron Law of IT Projects” alongside the well known Bent Flyvbjerg. Alex’s research has been pivotal in the project data analytics movement.

This episode looks at the value of data and whether it is the new oil or the oil spill. It dives into the uniqueness trap that causes project overruns and failures, with insight into how AI can help overcome these challenges. Alex provokes our thought as to where we can see humans bring immense value, pulling lessons from lean car manufacturing.

• Listen to the episode

• Read our blog of the episode

One more thing

The AI image generation hype is still making waves across social media. The latest is turning yourselves into toy collectibles. We had to ride the wave.

That’s it for today!

See you soon,

James, Yoshi and Aaron—Project Flux