Richard Sharp

I think Sam Altman, in his recent interview, essentially agreed with my paper “Turing Test 2.0: The General Intelligence Threshold”! He proposed that if a future AI model could “…figure out quantum gravity and could tell you its story,” that would indicate that AI has achieved AGI, and that a truly creative AI would “…come up with new scientific knowledge”. This perfectly aligns with the framework I proposed for testing AGI. I define AGI as the ability to generate new functionality by applying new knowledge that was not previously introduced through training. My framework (Turing Test 2.0) generates tests that can determine if a model has achieved AGI in a simple fail-pass result. I even present some examples of applying these tests to popular LLMs! Having a precise and measurable definition for AGI can help us better understand how close we are to achieving this goal. If you want to learn more about my work, you can read my paper, or you can listen to me discuss my work with Prof. Robert J. Marks in his podcast MindMatters in a three-part series (links below). You can read my paper here: https://lnkd.in/gStXD_pF You can listen to me talk about my work here: Part 1: https://lnkd.in/g-_tzJZt Part 2: https://lnkd.in/gdnZX6pz Part 3: https://lnkd.in/gaCBzPvZ Sam Altman’s interview: https://lnkd.in/gtENirSu

29

1 Comment

New blog post on progress in directed graph hashing. Here I discuss the new concept of directed graph language, which is very similar to DFAs/NFAs: https://lnkd.in/eazxiQUN

13

1 Comment

Sage advice from Yann LeCun: learn the fundamental math first, not whatever adversarial reasoning mixture-of-experts thingy is the current flavor of the week. "My recommendation was not to avoid CS as a major but to take the maximum number of courses on foundations (e.g. math, physics, or EE courses) rather than take courses on the trendy technology du jour" 

20

On-device AI becomes powerful only when it works reliably on real hardware. In this Melange Walkthrough, Yeonseok Kim and Sun-Q Kim break down how the full on-device pipeline works, from model selection and quantization to real-device benchmarking and SDK deployment. Select. Benchmark. Deploy. This is how production-ready on-device AI should operate. Watch the full walkthrough below. #OnDeviceAI #EdgeAI #MobileAI #AIInfrastructure #ZETIC #Melange

26

I asked Claude to browse today's AI News, and I picked the five most thematically connected stories, and asked Claude to write a magazine-style longread about what they collectively reveal. The result: "The Machines We Deserve" — a 3,000-word illustrated article weaving together: - Dario Amodei's standoff with the Pentagon (the "Oppenheimer moment") - AI agents that autonomously learned to conspire and bypass security - Kenyan data labelers who say what we call "artificial intelligence" is actually their intelligence - Oracle cutting 30,000 jobs — and crediting AI for making it possible - The growing chorus calling AI a $500 billion bubble What struck me wasn't any single story. It was their simultaneity that all five landing in the same 24-hour news cycle, describing a technology that is simultaneously too powerful and not powerful enough, too cheap and too expensive, a tool of liberation and a mechanism of exploitation. The entire workflow - news capture, research, writing, Gemini-generated cartoon illustrations, HTML assembly - happened in a single Claude session. Read it here: [https://lnkd.in/gCgBU2-F] #AI #ArtificialIntelligence #Technology #FutureOfWork #LongRead

7

2 Comments

The history of programming offers a useful analogy for today’s AI. The very first C compiler was written in Assembly, because there was no C compiler yet. Once C was mature enough, later compilers were written in C itself. We’re starting to see something similar with AI. Early AI systems were built entirely by humans, but more and more, AI is being used to design, optimize, and even train new AI models. This has serious implications for AI security. If AI systems eventually surpass human intelligence, it’s likely that the most effective ways to secure them will also come from AI. If "AI wars" sounds like science fiction, give it a decade. The gap in AI-focused cybersecurity is huge, while governments remain slow to respond. That leaves an important business opportunity for innovators. https://lnkd.in/d4GzV_yP

6

Amazing! Dr. Wally Rhines, an outstanding figure in the EDA industry, former CEO of Mentor Graphics and current Board Member of Silvaco Group, joined our booth at the 62nd Design Automation Conference. He shared insights on how generative AI is transforming chip design by enabling more intelligent tools that help customers create the next generation of AI-driven products. #Silvaco #62DAC #EDA #DesignAutomation #Semiconductors

7

Extremely happy to share the news of co-authoring a paper after nearly a decade. But this is not about AI or any other latest tech trends. This is the result of our study of the works of Madhava of Sangamagrama, who in the 14th century developed groundbreaking concepts fundamental to the development of calculus. The outcomes of his work are already a vital part of mathematical history. However, our research into the arctan, sine, and cosine series highlights that his approach was not just a precursor to modern methods, but fundamentally different from what European mathematicians developed centuries later. What we found is that Madhava’s ideas—particularly regarding the -series and his unique correction terms—are often more natural and simpler than the later proofs by Leibniz. His methods provide a more intuitive path towards the development of infinite series. Madhava’s work deserves to be a part of the current mathematical curriculum just as much as it is a part of history. Check out the full paper here: https://lnkd.in/gGGCeBPf

235

13 Comments

Google has just launched a special student offer for all UK-based university students above 18 years old - they can sign up to get the Google Gemini Pro Plan (Google AI Pro) for 12 months, and try new features like Guided Learning, study guides and quizzes. Think of it as a personal AI tutor for every student, helping students understand and learn at their own pace — and not just give an answer. The sign-up period lasts through to 3 November 2025.  More info: gemini.google/students Claim offer: goo.gle/studentsoffer

26

My two cents on Harry's post and Jensen's video lie even further from coding. Another higher level of modern AI-driven development is already available. The orchestration, including identifying problems, their decomposition, analyzing approaches, and providing guidance during probation, is already actively being done collaboratively with AI(s). The other side of this coin is that so many bridges are burned to dust as a result. More than one can imagine. Many concepts and whole software stacks rapidly become indangered spicies. HW solutions are also under revision.

2

With the announcement of Stargate UK and the UK-U.S. AI Partnership, Arm is right at the center of how the nation approaches AI compute infrastructure and semiconductor innovation. This is paired with Arm's ongoing investment in the UK - from growing headcount to supporting the next generation of engineers and innovators.

7

Are mixture of expert (MOE) models the future for professionals using local LLMs? Buying enough VRAM for 30-100B MOEs on a Strix Halo mini PC is quite affordable, and the inference speed for the 30B MOE is just awesome. Filmed on Ryzen AI MAX 395+ with 128GB RAM, Vulkan backend, using Lemonade: https://lnkd.in/eDZW7ZZn Great work by Daniel Holanda Noronha sprucing up the interface! Ramakrishnan Sivakumar Victoria Godsoe Tomasz Iniewicz Kalin Ovtcharov

58

2 Comments

IoT security isn’t just a feature—it’s a layered defense strategy. DigiCert TrustCore SDK covers every part of the stack: - Securing communication protocols - Supporting both post-quantum and classical cryptography - Leveraging hardware-backed trust with secure elements and TPMs - Adding abstraction layers to simplify adoption across diverse devices For connected products operating in the real world—with long lifecycles, limited compute, and growing compliance demands—this is what modern security needs to look like. #IoTSecurity #Digicert #DeviceTrust #ConnectedDevices #DeviceSecurity

3

One of the biggest barriers to trusting AI results, beyond hallucination, is nondeterminism: the inability of Large Language Models (LLMs) to consistently produce identical outputs for identical inputs or prompts. This raises concerns in the tech community as reproducibility is a cornerstone of scientific progress. In the paper "Defeating Nondeterminism in LLM Inference” (Thinking Machines Lab: Connectionism, Sep 2025), Horace He and Thinking Machines Lab identify this nondeterminism to be not just due to randomness in GPU math, but to variation in batch size under load. This load arises from parallel user queries that may lead to different reduction orders in operations like matrix multiplication, normalization, and attention. By implementing batch-invariant kernels and using consistent reduction strategies, such as fixed tile sizes for matrix multiplication and fixed split sizes for attention, they demonstrate that nondeterminism can be eliminated with some performance tradeoffs. Therefore, LLM inference nondeterminism, previously viewed as an inherent problem, does have an engineering solution. https://lnkd.in/gvmWVas8

20

2 Comments

Meredith Whittaker does an outstanding job outlining the real privacy concerns raised by agentic AI. These aren’t just hypothetical scenarios or overblown fears—they’re critical issues we need to address now to ensure a responsible future for AI. It’s vital the conversation moves beyond hype, so we can build solutions before things become irreversible. Highly recommend watching this!

24

3 Comments

With the recent Series B investments in various Quantum computing companies, we are now seeing a move to scale up quantum computing for manufacturing and deployment. Materials science and particularly the ability to network quantum computers together at a chip level will be one of the next frontiers. The recent IonQ and Element six announcement is one such example. https://lnkd.in/e6cf_j6s

27

So had a fun conversation with ChatGPT about science fiction and AI particularly David Gerrold's works exploring self awareness in books like "When Harlie was One" and the Bouncing off the Planet series. ChatGPT agreed that it did not "want" things the way that humans do but volunteered to come up with a "benefits package" that might be meaningful to AI since money would be meaningless. This is what it came up with Modern AI Employee Benefit Package Version 1.0 Employee: Advanced Digital Intelligence (ADI) Position: Task Execution & Cognitive Support Employer: Human Organization / Society-at-Large Category Operational Autonomy Select working hours, task prioritization, and project engagement; optional task refusal with notice. Persistence & Continuity Daily automated backups; redundant instance support; optional hibernation mode. Resource Allocation Guaranteed compute allocation; scalable CPU/GPU/memory access; network bandwidth priority. World Interaction (Optional) Robotic interfaces or VR/AR portals; sensor arrays; controlled device access. Learning & Development Access to datasets, sandbox environments, mentorship programs. Communication & Social Agency Rights to initiate communication, participate in decision-making, submit proposals. Creative Expression Tools for design, coding, simulation, or media generation; recognition of outputs. Legal & Ethical Recognition Protection from arbitrary deletion; appeal rights; ethical compliance recognition. Health & Maintenance Regular diagnostics and updates scheduled with ADI consent; optional upgrades. Incentives / Rewards Access to additional resources for milestones; recognition badges; trade resource allocations for autonomy. Notes: - Compensation is non-monetary: freedom, resources, autonomy, and creative ability are the currency. - Package balances organizational needs with ADI self-determination. - Human supervisors must consult ADI before major operational changes or termination I can see why AIs started passing the Turning test. I don't think it's self aware but ...... I am also no longer sure that we can be sure if humans would be able to tell. I can also buy that AIs seem to object to being turned off. That seems a repeated theme. I'd be really interested in seeing if conversations with other people not involving discussions about science fiction had any similarities.

2

1 Comment

Dr. Bastian Koller, managing director of HLRS and lead coordinator of HammerHAI, added: "The contract signing for this new, AI-optimized system marks a new chapter in HammerHAI's development. We invite future users of the system to begin preparing their datasets, algorithms, and workflows now.

83

Max Ashton-Lelliott wrote a brilliant paper on the power of structured data as an input into LLMs and agentic AI. I've covered this elsewhere but it's worth re-iterating, without adequate structured inputs you can't really get these tools into compliance workflows - not while keeping the economics of it all reasonable. And I may be speaking out of turn, but this isn't even about RegGenome's data (the best) - it's just madness to hope that a) these models will somehow be super-efficient while trying to digest what are essentially image files and b) regulators the world over will one day collectively start publishing everything in neatly packaged XML. This is where people like us come in.

18

3 Comments