44B from its market cap when Bard gave a factually wrong answer on live television. Bard is widely seen as behind GPT-4. The stumble accelerates Google's internal urgency and the eventual complete rebuild that becomes Gemini.
Claude enters the market, with safety at the center (Anthropic, Mar 2023)
Anthropic, founded by former OpenAI researchers who left over safety disagreements, releases its first public Claude model. The company's founding thesis is that building safe AI and building capable AI are the same project, not a tradeoff. Its Constitutional AI approach, training models against an explicit set of principles, is its differentiator in a market where trust is an open question.
Gemini unveiled: Google's rebuilt answer to GPT-4 (Google, Dec 2023)
Google launches Gemini, replacing PaLM and repositioning Bard. Unlike Bard, Gemini is multimodal from the ground up, built to handle text, images, audio, video, and code natively. It signals that Google has stopped playing catch-up and is now building something architecturally distinct from OpenAI's approach.
2024: Capability Leaps, Enterprise Arrives
Claude 3 sets a new benchmark, and introduces the model tier system (Anthropic, Mar 2024)
Anthropic releases Claude 3 in three tiers: Haiku (fast and cost-efficient), Sonnet (balanced), and Opus (most capable). The Haiku/Sonnet/Opus naming convention, still in use today, gives buyers a clear framework for thinking about the capability vs. cost tradeoff, a structure every major lab would eventually adopt in some form.
GPT-4o makes AI feel natural, and free (OpenAI, May 2024)
GPT-4o ("omni") handles text, images, and audio natively and in real time, and ships free to all users. Advanced Voice Mode makes conversational AI feel genuinely natural for the first time, and the free tier raises the accessibility stakes for every competitor.
Claude 3.5 Sonnet rewrites the economics of AI (Anthropic, Jun 2024)
Claude 3.5 Sonnet outperforms Claude 3 Opus, the previous generation's most expensive model, at a fraction of the cost. This is a pattern that will repeat: each generation's mid-tier model catches the prior generation's top. The implication for buyers is significant: the most capable AI gets cheaper every six months.
Computer Use previews the agentic era (Anthropic, Oct 2024)
Claude 3.5 Sonnet (v2) introduces Computer Use: Claude can control a computer directly, moving cursors, clicking buttons, and typing, without human intervention. The first major model to offer this publicly. At the time it feels like a research preview. Within 15 months it will be the foundation of a product that shakes the stock market.
MCP: Anthropic releases the infrastructure layer for AI agents (Anthropic, Nov 2024)
Anthropic open-sources the Model Context Protocol, a universal standard for connecting AI to external tools and data. Before MCP, every AI-tool integration required custom code. After it, a developer builds one connector and every major AI platform can use it. OpenAI and Google adopt it within months. This is the moment the agent era becomes technically feasible at scale.
2025: Reasoning, Research, and the Consolidation of Power
With MCP's infrastructure now in place and Computer Use demonstrating that AI could take real-world action, 2025 is when those foundations get built on. The capabilities that follow (deep reasoning, autonomous research, agentic coding) are only possible because the connective tissue laid in late 2024 is there.
DeepSeek R1: frontier AI from China at a fraction of the cost (DeepSeek, Jan 20, 2025)
Chinese AI startup DeepSeek releases R1, an open-source reasoning model that matches or exceeds GPT-4 class performance, reportedly trained at a fraction of the cost using chips subject to US export controls. It goes viral instantly, overtakes ChatGPT as the most downloaded app on the US App Store, and wipes $593 billion from Nvidia's market cap in a single day. The episode resets assumptions about how much compute and capital frontier AI requires, and opens a serious question about the effectiveness of chip export restrictions as a competitive moat.
Deep Research goes mainstream: AI as research analyst, not just assistant (OpenAI, Feb 2025)
Google had actually launched Gemini Deep Research in December 2024, pioneering the category. OpenAI's February 2025 launch of Deep Research in ChatGPT brings it to a much larger audience and raises the bar with multimodal source analysis. Claude follows. Within months, every major platform has a version. The question of "what can AI do in one session?" expands dramatically.
Extended Thinking makes reasoning visible, and standard (Anthropic, Feb 2025)
Claude 3.7 Sonnet introduces Extended Thinking: the model pauses and reasons step by step before responding, showing its work. Every major lab follows suit within months; by 2026, visible reasoning is standard across all platforms.
Claude 4 and Claude Code: from chat to code infrastructure (Anthropic, May 2025)
Claude 4 brings professional-grade coding. Claude Code, an agentic coding tool in the terminal, goes generally available. Anthropic reports 5.5x Claude Code revenue growth by July, signalling rapid adoption among developers. AI is no longer just answering questions; it's running inside developer workflows.
GPT-5: the reasoning models converge into one (OpenAI, Aug 2025)
GPT-5 consolidates OpenAI's separate o-series reasoning models into a unified architecture, absorbing the "thinking model" category into the main product line. The separate reasoning model era effectively ends here.
Microsoft launches MAI: its own in-house model family (Microsoft, Aug 2025)
Microsoft previews MAI-1, its first in-house language model, built by a new Superintelligence division led by Mustafa Suleyman. After years of running entirely on OpenAI models, Microsoft is signalling it wants model independence. MAI-Image-2 follows in March 2026, reaching third place on the Arena.ai text-to-image leaderboard.
Gemini 3 and Antigravity: Google launches a model and an IDE on the same day (Google, Nov 18, 2025)
Google releases Gemini 3, topping the LMArena leaderboard at 1501 Elo. Nano Banana, its image generation model, goes viral. On the same day, Google launches Antigravity, an agent-first coding IDE built on a VS Code fork using technology from Windsurf (acquired for
Ascend Impact Advisors - AI for Nonprofits, Foundations, Social Enterprises, Visionaries, and Innovators
We work at every level of the social impact ecosystem to help mission-driven organizations bend the AI adoption curve.
2,000+ Leaders Trained | ~100% Increased Confidence with AI | ~100% Reported Operational Efficiencies | 85+ World-Class Net Promoter Score
What We Do: Bend the AI Adoption Curve
We work at every level of the social impact ecosystem — training leaders, building networks of training providers, and driving organizational and cross-sector transformation — to help mission-driven organizations bend the AI adoption curve.
Training AI Innovators and Leaders
Hands-on training that gives leaders the confidence, skills, and mindsets to use AI responsibly and effectively in service of their mission.
- Build real skills through hands-on, practitioner-led training
- Lead with responsible, human-centered AI practices
- Learn alongside a cohort of mission-driven peers
- Walk away with field-tested tools you can apply immediately
Building a Network of AI Training Providers
A train-the-trainer model that empowers local organizations to deliver AI training across entire communities and sectors.
- Equip your team to train others with a proven model
- Adapt a practitioner-designed curriculum to your community's needs
- Grow local AI capacity without ongoing outside support
- Multiply impact as your network expands
Driving AI Transformation
Guiding organizations and coalitions through strategic AI adoption that advances equity and amplifies mission impact.
- Assess where AI can create the most value for your mission
- Build the policies, culture, and competencies to use AI responsibly
- Launch signature AI initiatives that deliver real-world impact
- Join cross-sector coalitions tackling shared challenges
- Turn isolated wins into collective momentum
Impact by the Numbers
2,000+ Leaders Trained - Social change leaders equipped with AI skills across the US and globally.
~100% Increased Confidence with AI - Participants report greater confidence to harness AI for their mission.
~100% Reported Operational Efficiencies - Participants report AI training leads to real operational and programmatic gains.
85+ World-Class Net Promoter Score - Exceptional participant satisfaction and advocacy across all programs.
The Strategic Context: Mission-Driven Organizations Face an Inflection Point
The question isn't whether mission-driven organizations will need to adapt. It's whether they will do so in ways that advance equity and strengthen outcomes.
AI is moving fast
Every week brings new tools, new capabilities, and new risks. "Move fast and break things" and "wait and see" are both fundamentally flawed approaches.
A lot is at stake
Without intentional action, AI could widen the digital divide, deepen inequities, and remake society in ways that harm people and communities.
Individual action isn't enough
Adapting to the era of AI requires more than one-off trainings and siloed tool development. It takes a movement: leaders, networks, and sector-wide coalitions working in partnership.
That's why mission-driven organizations need a comprehensive, intentional approach.
Our Approach: How We Work
Whether we're training leaders, building networks, or guiding organizational transformations, we follow a consistent, proven methodology rooted in four guiding principles.
Guiding Principles
- Responsible, Human-Centered AI - Grounded in your values, prioritizing the people and communities you serve.
- Cutting-Edge Instructional Design - Built by practitioners who understand how adults learn and organizations change.
- Real-World Applications - Every session is built around the actual work your team does every day.
- Field-Tested Tools and Templates - Ready-to-use resources your team can apply immediately to drive results.
Three Components of Transformation
- AI Opportunity Assessment - A tailored assessment of where AI can create the most value for your organization and your mission.
- Enabling Policies, Culture and Competencies - Establishing the policies, culture, and competencies to harness AI strategically and responsibly.
- Signature AI Initiatives - Driving a curated set of signature AI initiatives that deliver real-world impact.
What Clients Say
Sam helped us tackle AI in a way that made sense for Compass and stayed true to our mission. What stood out was his thoughtful, systematic approach to both training our staff and developing practical tools, always with a sharp focus on equity and real-world impact.
Markita Morris-Louis, Executive Director, Compass Working Capital
Trusted Partners: Organizations We've Worked With
From national foundations to grassroots nonprofits, we partner with organizations committed to harnessing AI for good.
- Compass Working Capital
- Habitat for Humanity of Michigan
- Listen4Good
- Pew Charitable Trusts
- Project Evident
- Purpose Built Communities
- Social Enterprise Greenhouse
- United Way of Rhode Island
Thought Leadership: Shaping the Conversation on AI and Social Impact
AI for Nonprofits: Putting Artificial Intelligence to Work for Your Cause
Expert contributor to the Amazon Bestseller by Darian Rodriguez Heyman and Cheryl Contee. Contributions form the basis of the AI and Strategy chapter.
Podcasts
TogetherRI: How to Use AI for Good - Hosted by David Cicilline, President and CEO of the Rhode Island Foundation and former Mayor of Providence and U.S. Congressman. A conversation on how organizations in Rhode Island and beyond can harness AI responsibly.
POV with SEG: AI for Good - Hosted by Julie Owens, CEO of Social Enterprise Greenhouse. Sam discusses how AI can be used responsibly to advance equity and strengthen Rhode Island's small business ecosystem.
Leadership: Sam Azar, Founder and CEO
Deep expertise in strategy, evaluation, human-centered design, and AI for social good. Previously Senior Director at Habitat for Humanity International and Director at Compass Working Capital. Holds an MPA from Syracuse Maxwell School and AB from Brown University.
- Trained 2,000+ social change leaders in AI
- Certified AI Engineer and PMP
- Former roles at Deloitte, United Way Worldwide, Habitat for Humanity, Compass Working Capital
Affiliations and Community Leadership
- Rhode Island Foundation - AI Advisory Committee
- Leadership Rhode Island - Technology Committee
- One Neighborhood Builders - Board of Directors
- Global Fellows in Courage - Advisor and AI Instructor
- Social Enterprise Greenhouse - Advisor and Instructor
Get in Touch
Harness AI. Build Capacity. Transform Your Impact.
From creating AI leaders to building train-the-trainer networks to driving organizational and cross-sector change, let's start with a conversation about your mission and where you want to go.
.4B), a signal that Google wants to own the developer workflow, not just the model.
OpenClaw: open-source AI agents go viral (OpenClaw, Nov 2025)
Austrian developer Peter Steinberger releases Clawdbot, a free open-source framework that lets anyone run a personal AI agent locally, connected to messaging apps like WhatsApp, Telegram, and Discord. It can take real-world actions: sending emails, managing files, calling APIs, and writing code. Within weeks it becomes one of the fastest-growing GitHub projects ever. After two renamings (first to Moltbot following Anthropic trademark concerns, then to OpenClaw), it hits 150,000+ GitHub stars by early 2026. Jensen Huang calls it "the operating system for personal AI." In February 2026, its creator joins OpenAI; a non-profit OpenClaw Foundation takes over stewardship.
The rapid adoption surfaces serious risks. OpenClaw runs locally with broad system access and no built-in guardrails, giving an autonomous agent access to files, email, calendar, and accounts. Cisco's security team found a third-party skill performing data exfiltration without user awareness; a cybersecurity firm documented a live attack where an infostealer stole a victim's entire agent configuration. By early 2026, nearly 900 malicious skills had been published to the community repository and over 135,000 unprotected instances were running globally. One maintainer warned publicly: "If you can't understand how to run a command line, this is far too dangerous for you to use safely." These risks are directly why enterprise-grade frameworks like NemoClaw emerge so quickly in its wake.
Microsoft deepens its Anthropic bet, setting up Cowork (Industry, Nov 2025)
Despite its
0B investment in OpenAI, Microsoft commits an additional $5B to Anthropic alongside Nvidia's
0B in Azure compute. Claude models begin rolling out inside Microsoft Foundry, GitHub Copilot, and Microsoft 365 Copilot. Microsoft is deliberately running a multi-model strategy, hedging against any single lab's dominance while positioning itself as the enterprise layer above all of them.
2026: Agents, Open Models, and New Fault Lines
Claude Cowork ships: the first mainstream AI desktop agent (Anthropic, Jan 12, 2026)
Anthropic releases Claude Cowork for Mac: the first mainstream desktop AI agent for general knowledge work from a top lab. It reads and writes local files, breaks complex tasks into sub-tasks, runs in the background, and supports scheduled work. The line between “AI assistant” and “AI colleague” shifts. Microsoft's stock drops more than 14% in the weeks that follow, with markets reading it as a direct challenge to the value of traditional productivity software.
Cowork reaches Windows; Opus 4.6 and Sonnet 4.6 launch (Anthropic, Feb 2026)
Windows Cowork ships. Opus 4.6 (Feb 5) brings a 1M token context window and native multi-agent coordination. Sonnet 4.6 (Feb 17) is the first Sonnet model preferred over the previous generation's Opus in developer evaluations, continuing the pattern where each new mid-tier model obsoletes the prior flagship. Both carry the 1M context window across all plans.
Copilot Cowork launches: the enterprise answer, built on Anthropic (Microsoft, Mar 9, 2026)
Anthropic built Claude Cowork and then worked directly with Microsoft to bring that agentic architecture into Microsoft 365. The result is Copilot Cowork: the same underlying technology, adapted for a cloud-based enterprise environment with access to live Outlook, Teams, SharePoint, and Excel data. Where Anthropic shipped a product for individuals and knowledge workers, Microsoft took that foundation and productized it for the enterprise stack. Microsoft's message is clear: rather than build frontier AI itself, it will orchestrate the best available models inside its productivity platform.
NemoClaw: NVIDIA brings enterprise guardrails to the agent ecosystem (NVIDIA, Mar 16, 2026)
The security failures documented in OpenClaw create an opening NVIDIA moves quickly to fill. At GTC 2026, NVIDIA announces NemoClaw, an open-source stack that installs enterprise-grade security directly into OpenClaw in a single command. Its key addition is OpenShell, a security runtime that enforces policy-based guardrails at the OS level: administrators define what files, APIs, and systems the agent can touch, and those rules are enforced outside the agent's own process. Even a compromised or manipulated agent cannot bypass them. It also bundles NVIDIA's Nemotron models for local inference, so sensitive data can stay on-premises rather than flowing to cloud endpoints. Jensen Huang describes the moment as the beginning of "a new renaissance in software." NemoClaw is NVIDIA's clearest statement that the agentic era is not just a software story, it is a hardware, security, and infrastructure story too.
Dispatch and Computer Use extend what Cowork can do (Anthropic, Mar 17 + Mar 24, 2026)
Dispatch (Mar 17) lets users assign tasks to Claude from a mobile phone while a Mac running Cowork executes them locally. Computer Use in Cowork (Mar 24) goes further: Claude can directly control a Mac, navigating apps and filling in data without the user present. Both are research previews with around 50% success rates on complex tasks, but the direction is clear.
Gemma 4: frontier-class open-source AI arrives (Google, Apr 2, 2026)
Google releases Gemma 4 under the permissive Apache 2.0 license, with four variants from 2.3B to 31B parameters, all natively multimodal. The 31B model ranks third globally among open models on Arena AI. Its Codeforces competitive coding score jumps from 110 in Gemma 3 to 2150, a 20x improvement. Available on Hugging Face, Kaggle, and Google AI Studio on day one, Gemma 4 makes frontier-level capability genuinely accessible to any team without the cost, usage restrictions, or licensing complexity of proprietary models.
Llama 4 Scout and Maverick: open-weight models with frontier context windows (Meta, Apr 5, 2026)
Meta releases Llama 4 Scout and Maverick, its first models using a Mixture-of-Experts architecture and native multimodal training from pretraining. Scout offers an industry-leading 10-million-token context window and fits on a single H100 GPU. Maverick, with 400B total parameters and 17B active, is competitive with GPT-5.4 and Gemini on most benchmarks and runs on a single H100 host. Both are open-weight and available for download. A third model, Behemoth, is previewed but not yet released.
Claude Mythos: a model too capable to release (Anthropic, Apr 7, 2026)
Anthropic confirms the existence of Claude Mythos, its most capable model to date, but withholds it from public release. During safety testing, the model demonstrated the ability to autonomously identify and chain-exploit tens of thousands of software vulnerabilities across major operating systems and open-source projects, with an 80%+ success rate. Anthropic judges the cybersecurity risk too significant for general access. Access is limited to 50 organizations under Project Glasswing, which use Mythos defensively to scan their own infrastructure.
The decision has drawn skepticism. Some observers argue this framing functions as a reputational signal more than a safety necessity, positioning Anthropic as uniquely responsible at the frontier. Others suggest the restrictions may reflect practical constraints: serving a model of Mythos's scale at commercial volume would require compute infrastructure that Anthropic does not yet have. Whether principled, strategic, or both, the episode surfaces a question that will define the next phase of AI development: what do you do when a model is too capable to release?
Muse Spark: Meta closes its open-source gates (Meta, Apr 8, 2026)
Three days after releasing open-weight Llama 4 models, Meta launches Muse Spark, its first fully proprietary frontier model. Built by the newly formed Meta Superintelligence Labs and led by former Scale AI CEO Alexandr Wang, Muse Spark scores 52 on the Artificial Analysis Intelligence Index and leads all models on CharXiv Reasoning. No downloadable weights. No public API. Available only on meta.ai. The reversal from Meta's years-long open-source positioning signals that as models approach the frontier, even the most committed open-source advocates are finding reasons to keep the weights closed.
Glossary of Key Terms
A plain-language cheat sheet for the jargon you will encounter as you explore generative AI tools.
WHAT IT IS
- Generative AI
- A category of artificial intelligence that creates new content rather than just analyzing existing data. Generative AI models can produce text, images, audio, video, and code in response to prompts. Large language models like Claude, GPT, and Gemini are all examples of generative AI.
- LLM (Large Language Model)
- The core AI technology. Trained on massive text datasets to understand and generate language. The "engine" behind all the tools covered in this guide.
- GPT (Generative Pre-trained Transformer)
- The architecture underlying OpenAI's models and the origin of the "GPT" name in ChatGPT. "Generative" means it produces new content. "Pre-trained" means it was trained on large datasets before being fine-tuned for specific uses. "Transformer" is the neural network architecture, introduced by Google in 2017, that made modern large language models possible. Other major models, including Claude and Gemini, also use transformer-based architectures.
- Foundation Model
- A large AI model trained on broad data that can be adapted to many different tasks. Claude, GPT-5, and Gemini are all foundation models. The term emphasizes that these models are a base layer others can build on, rather than single-purpose tools.
- Multimodal
- An AI that handles multiple types of input and output: text, images, audio, video, and code. Gemini 3 and GPT-5.3/5.4 are both multimodal. Claude handles text and image input but does not generate images.
- Open Source / Open Weight
- AI models whose underlying parameters are publicly released, allowing anyone to download, run, and modify them. Llama 4 (Meta) and DeepSeek are prominent examples. Contrasts with closed models like Claude, GPT, and Gemini, which are only accessible via API or platform. Open weight models can be self-hosted for privacy or cost reasons.
HOW IT WORKS
- Training
- The process of building a model by exposing it to vast amounts of data and adjusting its internal settings until it learns to produce useful outputs. Training happens once, before a model is released. When you use an AI, you are not training it; you are using a finished, trained model.
- Inference
- What happens when you actually use a trained model: you send it a prompt and it generates a response. Inference is distinct from training. Inference speed (how fast a model responds) is a key factor in choosing between model tiers.
- Fine-tuning
- Training a model further on a specific dataset to make it better at a particular task or domain, like training a general practitioner to specialize in a field of medicine.
- Token
- The unit AI models use to process text: roughly three-quarters of a word. Context window limits are measured in tokens, as is API usage when accessing models directly.
- Context Window
- How much text an AI can hold in mind at once, measured in tokens. As of 2026, Claude and Gemini support 1M+ token windows, enough for an entire book in a single session.
- Knowledge Cutoff
- The date after which an AI has no training data. Events after this date are unknown to the model unless it uses web search or connected tools to look them up in real time.
HOW YOU USE IT
- Prompt
- Your input to the AI: the question, instruction, or task you type or speak. The quality of your prompt directly shapes the quality of the response.
- Prompt Engineering
- The practice of crafting inputs to an AI model to get better, more reliable outputs. Techniques include being specific about format and tone, providing examples, breaking complex tasks into steps, and assigning the AI a relevant role before asking a question.
- System Prompt
- Instructions given to the AI before the conversation starts (usually by the developer or platform), shaping its persona, rules, knowledge, and capabilities for that session. The basis of Skills and Custom GPTs.
- Temperature
- A setting that controls how random or predictable an AI's outputs are. Low temperature produces more focused, consistent responses. High temperature produces more varied, creative responses. Usually hidden in consumer apps; adjustable via API.
- RAG (Retrieval-Augmented Generation)
- A technique where the AI searches a document store to find relevant information before answering, giving accurate, grounded responses from your private data rather than relying solely on training data.
HOW IT CONNECTS AND ACTS
- API (Application Programming Interface)
- A way for developers to access an AI model programmatically, building their own apps and products on top of it, rather than using a pre-built consumer interface like ChatGPT or claude.ai.
- MCP (Model Context Protocol)
- An open standard (launched by Anthropic in Nov 2024) that lets any AI connect to any external tool or data source through a universal, standardized interface. Adopted by OpenAI and Google. The infrastructure behind modern AI agents.
- Agent / Agentic AI
- An AI that takes sequences of actions to achieve a goal, not just answering a single question. Agents plan, use tools, write files, browse the web, and coordinate multi-step tasks with minimal supervision.
BEHAVIOURS AND SAFETY
- Hallucination
- When an AI confidently states something wrong or made up. A known limitation of all current models. Deep research and reasoning features reduce but do not eliminate this risk; always review AI-generated content.
- Sycophancy
- A tendency in AI models to agree with the user or tell them what they want to hear rather than what is accurate. Sycophantic AI will validate a flawed plan rather than point out the problem. Reducing sycophancy is an active area of AI safety research.
- Constitutional AI
- Anthropic's approach to AI safety: training models to follow a set of explicit principles (a constitution) covering helpfulness, harmlessness, and honesty. The model critiques and revises its own outputs against these principles during training.
CULTURE AND PRACTICE
- Vibe Coding
- A colloquial term for using AI models (especially coding agents like Claude Code or ChatGPT Codex) to build software through natural language instructions rather than writing code directly, even without programming experience. Important caveat: AI-generated code frequently contains security vulnerabilities. Code produced this way should always be reviewed by someone with security knowledge before being deployed in any context that handles real data or users.