#ai secure model

[willcodehtmlforfood]

Hugging Face, the GitHub of AI, hosted code that backdoored user devices | Ars Technica

"Code uploaded to AI developer platform Hugging Face covertly installed backdoors and other types of malware on end-user machines, researchers from security firm JFrog said Thursday in a report that’s a likely harbinger of what’s to come.

In all, JFrog researchers said, they found roughly 100 submissions that performed hidden and unwanted actions when they were downloaded and loaded onto an end-user device. Most of the flagged machine learning models—all of which went undetected by Hugging Face—appeared to be benign proofs of concept uploaded by researchers or curious users. JFrog researchers said in an email that 10 of them were “truly malicious” in that they performed actions that actually compromised the users’ security when loaded."

[atcuality1]

Simplify Transactions and Boost Efficiency with Our Cash Collection Application

Manual cash collection can lead to inefficiencies and increased risks for businesses. Our cash collection application provides a streamlined solution, tailored to support all business sizes in managing cash effortlessly. Key features include automated invoicing, multi-channel payment options, and comprehensive analytics, all of which simplify the payment process and enhance transparency. The application is designed with a focus on usability and security, ensuring that every transaction is traceable and error-free. With real-time insights and customizable settings, you can adapt the application to align with your business needs. Its robust reporting functions give you a bird’s eye view of financial performance, helping you make data-driven decisions. Move beyond traditional, error-prone cash handling methods and step into the future with a digital approach. With our cash collection application, optimize cash flow and enjoy better financial control at every level of your organization.

[jcmarchi]

Qwen2-Math: A new era for AI maths whizzes

New Post has been published on https://thedigitalinsider.com/qwen2-math-a-new-era-for-ai-maths-whizzes/

Qwen2-Math: A new era for AI maths whizzes

.pp-multiple-authors-boxes-wrapper display:none; img width:100%;

Alibaba Cloud’s Qwen team has unveiled Qwen2-Math, a series of large language models specifically designed to tackle complex mathematical problems.

These new models – built upon the existing Qwen2 foundation – demonstrate remarkable proficiency in solving arithmetic and mathematical challenges, and outperform former industry leaders.

The Qwen team crafted Qwen2-Math using a vast and diverse Mathematics-specific Corpus. This corpus comprises a rich tapestry of high-quality resources, including web texts, books, code, exam questions, and synthetic data generated by Qwen2 itself.

Rigorous evaluation on both English and Chinese mathematical benchmarks – including GSM8K, Math, MMLU-STEM, CMATH, and GaoKao Math – revealed the exceptional capabilities of Qwen2-Math. Notably, the flagship model, Qwen2-Math-72B-Instruct, surpassed the performance of proprietary models such as GPT-4o and Claude 3.5 in various mathematical tasks.

“Qwen2-Math-Instruct achieves the best performance among models of the same size, with RM@8 outperforming Maj@8, particularly in the 1.5B and 7B models,” the Qwen team noted.

This superior performance is attributed to the effective implementation of a math-specific reward model during the development process.

Further showcasing its prowess, Qwen2-Math demonstrated impressive results in challenging mathematical competitions like the American Invitational Mathematics Examination (AIME) 2024 and the American Mathematics Contest (AMC) 2023.

To ensure the model’s integrity and prevent contamination, the Qwen team implemented robust decontamination methods during both the pre-training and post-training phases. This rigorous approach involved removing duplicate samples and identifying overlaps with test sets to maintain the model’s accuracy and reliability.

Looking ahead, the Qwen team plans to expand Qwen2-Math’s capabilities beyond English, with bilingual and multilingual models in the pipeline.  This commitment to inclusivity aims to make advanced mathematical problem-solving accessible to a global audience.

“We will continue to enhance our models’ ability to solve complex and challenging mathematical problems,” affirmed the Qwen team.

You can find the Qwen2 models on Hugging Face here.

See also: Paige and Microsoft unveil next-gen AI models for cancer diagnosis

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is co-located with other leading events including Intelligent Automation Conference, BlockX, Digital Transformation Week, and Cyber Security & Cloud Expo.

Explore other upcoming enterprise technology events and webinars powered by TechForge here.

Tags: ai, alibaba cloud, artificial intelligence, maths, models, qwen, qwen2, qwen2-math

[filehulk]

What is WormGPT?

Artificial intelligence (AI) tools are expected to transform the workplace by automating everyday tasks, increasing productivity for everyone. However, AI can also be misused for illegal activities, as highlighted by the new WormGPT system. What is WormGPT? WormGPT is a harmful AI tool designed for cybercriminal activities. It is based on the GPTJ language model, developed by OpenAI, and was…

[bitcoinversus]

Scale AI Unveils Defense Llama for U.S. National Security

Scale AI has unveiled Defense Llama, a Large Language Model (LLM) developed in collaboration with Meta and defense experts, tailored specifically for U.S. national security applications. This model is accessible exclusively within controlled U.S. government environments through Scale Donovan, enabling military and national security personnel to leverage generative AI for tasks such as military…

[jamaicahomescom]

The Future of Real Estate in Jamaica: AI, Big Data, and Cybersecurity Shaping Tomorrow’s Market

[ai-innova7ions]

Video Automatically Generated by Faceless.Video:

Revolutionize Your Data Security with AI Models!

Small language models are revolutionizing the tech landscape by providing a more efficient alternative to larger counterparts. Their ability to operate on modest hardware means they can run locally, making them perfect for industries like healthcare and finance where data privacy is crucial.

These models enable fast and secure processing of sensitive information, transforming how businesses manage data. By addressing privacy concerns without compromising performance, small language models are paving the way for innovative solutions in various sectors. Discover how these advancements impact our world today!

#SmallLanguageModels

#DataPrivacy

[phantomrose96]

If anyone wants to know why every tech company in the world right now is clamoring for AI like drowned rats scrabbling to board a ship, I decided to make a post to explain what's happening.

(Disclaimer to start: I'm a software engineer who's been employed full time since 2018. I am not a historian nor an overconfident Youtube essayist, so this post is my working knowledge of what I see around me and the logical bridges between pieces.)

Okay anyway. The explanation starts further back than what's going on now. I'm gonna start with the year 2000. The Dot Com Bubble just spectacularly burst. The model of "we get the users first, we learn how to profit off them later" went out in a no-money-having bang (remember this, it will be relevant later). A lot of money was lost. A lot of people ended up out of a job. A lot of startup companies went under. Investors left with a sour taste in their mouth and, in general, investment in the internet stayed pretty cooled for that decade. This was, in my opinion, very good for the internet as it was an era not suffocating under the grip of mega-corporation oligarchs and was, instead, filled with Club Penguin and I Can Haz Cheezburger websites.

Then around the 2010-2012 years, a few things happened. Interest rates got low, and then lower. Facebook got huge. The iPhone took off. And suddenly there was a huge new potential market of internet users and phone-havers, and the cheap money was available to start backing new tech startup companies trying to hop on this opportunity. Companies like Uber, Netflix, and Amazon either started in this time, or hit their ramp-up in these years by shifting focus to the internet and apps.

Now, every start-up tech company dreaming of being the next big thing has one thing in common: they need to start off by getting themselves massively in debt. Because before you can turn a profit you need to first spend money on employees and spend money on equipment and spend money on data centers and spend money on advertising and spend money on scale and and and

But also, everyone wants to be on the ship for The Next Big Thing that takes off to the moon.

So there is a mutual interest between new tech companies, and venture capitalists who are willing to invest $$$ into said new tech companies. Because if the venture capitalists can identify a prize pig and get in early, that money could come back to them 100-fold or 1,000-fold. In fact it hardly matters if they invest in 10 or 20 total bust projects along the way to find that unicorn.

But also, becoming profitable takes time. And that might mean being in debt for a long long time before that rocket ship takes off to make everyone onboard a gazzilionaire.

But luckily, for tech startup bros and venture capitalists, being in debt in the 2010's was cheap, and it only got cheaper between 2010 and 2020. If people could secure loans for ~3% or 4% annual interest, well then a $100,000 loan only really costs $3,000 of interest a year to keep afloat. And if inflation is higher than that or at least similar, you're still beating the system.

So from 2010 through early 2022, times were good for tech companies. Startups could take off with massive growth, showing massive potential for something, and venture capitalists would throw infinite money at them in the hopes of pegging just one winner who will take off. And supporting the struggling investments or the long-haulers remained pretty cheap to keep funding.

You hear constantly about "Such and such app has 10-bazillion users gained over the last 10 years and has never once been profitable", yet the thing keeps chugging along because the investors backing it aren't stressed about the immediate future, and are still banking on that "eventually" when it learns how to really monetize its users and turn that profit.

The pandemic in 2020 took a magnifying-glass-in-the-sun effect to this, as EVERYTHING was forcibly turned online which pumped a ton of money and workers into tech investment. Simultaneously, money got really REALLY cheap, bottoming out with historic lows for interest rates.

Then the tide changed with the massive inflation that struck late 2021. Because this all-gas no-brakes state of things was also contributing to off-the-rails inflation (along with your standard-fare greedflation and price gouging, given the extremely convenient excuses of pandemic hardships and supply chain issues). The federal reserve whipped out interest rate hikes to try to curb this huge inflation, which is like a fire extinguisher dousing and suffocating your really-cool, actively-on-fire party where everyone else is burning but you're in the pool. And then they did this more, and then more. And the financial climate followed suit. And suddenly money was not cheap anymore, and new loans became expensive, because loans that used to compound at 2% a year are now compounding at 7 or 8% which, in the language of compounding, is a HUGE difference. A $100,000 loan at a 2% interest rate, if not repaid a single cent in 10 years, accrues to $121,899. A $100,000 loan at an 8% interest rate, if not repaid a single cent in 10 years, more than doubles to $215,892.

Now it is scary and risky to throw money at "could eventually be profitable" tech companies. Now investors are watching companies burn through their current funding and, when the companies come back asking for more, investors are tightening their coin purses instead. The bill is coming due. The free money is drying up and companies are under compounding pressure to produce a profit for their waiting investors who are now done waiting.

You get enshittification. You get quality going down and price going up. You get "now that you're a captive audience here, we're forcing ads or we're forcing subscriptions on you." Don't get me wrong, the plan was ALWAYS to monetize the users. It's just that it's come earlier than expected, with way more feet-to-the-fire than these companies were expecting. ESPECIALLY with Wall Street as the other factor in funding (public) companies, where Wall Street exhibits roughly the same temperament as a baby screaming crying upset that it's soiled its own diaper (maybe that's too mean a comparison to babies), and now companies are being put through the wringer for anything LESS than infinite growth that Wall Street demands of them.

Internal to the tech industry, you get MASSIVE wide-spread layoffs. You get an industry that used to be easy to land multiple job offers shriveling up and leaving recent graduates in a desperately awful situation where no company is hiring and the market is flooded with laid-off workers trying to get back on their feet.

Because those coin-purse-clutching investors DO love virtue-signaling efforts from companies that say "See! We're not being frivolous with your money! We only spend on the essentials." And this is true even for MASSIVE, PROFITABLE companies, because those companies' value is based on the Rich Person Feeling Graph (their stock) rather than the literal profit money. A company making a genuine gazillion dollars a year still tears through layoffs and freezes hiring and removes the free batteries from the printer room (totally not speaking from experience, surely) because the investors LOVE when you cut costs and take away employee perks. The "beer on tap, ping pong table in the common area" era of tech is drying up. And we're still unionless.

Never mind that last part.

And then in early 2023, AI (more specifically, Chat-GPT which is OpenAI's Large Language Model creation) tears its way into the tech scene with a meteor's amount of momentum. Here's Microsoft's prize pig, which it invested heavily in and is galivanting around the pig-show with, to the desperate jealousy and rapture of every other tech company and investor wishing it had that pig. And for the first time since the interest rate hikes, investors have dollar signs in their eyes, both venture capital and Wall Street alike. They're willing to restart the hose of money (even with the new risk) because this feels big enough for them to take the risk.

Now all these companies, who were in varying stages of sweating as their bill came due, or wringing their hands as their stock prices tanked, see a single glorious gold-plated rocket up out of here, the likes of which haven't been seen since the free money days. It's their ticket to buy time, and buy investors, and say "see THIS is what will wring money forth, finally, we promise, just let us show you."

To be clear, AI is NOT profitable yet. It's a money-sink. Perhaps a money-black-hole. But everyone in the space is so wowed by it that there is a wide-spread and powerful conviction that it will become profitable and earn its keep. (Let's be real, half of that profit "potential" is the promise of automating away jobs of pesky employees who peskily cost money.) It's a tech-space industrial revolution that will automate away skilled jobs, and getting in on the ground floor is the absolute best thing you can do to get your pie slice's worth.

It's the thing that will win investors back. It's the thing that will get the investment money coming in again (or, get it second-hand if the company can be the PROVIDER of something needed for AI, which other companies with venture-back will pay handsomely for). It's the thing companies are terrified of missing out on, lest it leave them utterly irrelevant in a future where not having AI-integration is like not having a mobile phone app for your company or not having a website.

So I guess to reiterate on my earlier point:

Drowned rats. Swimming to the one ship in sight.

[airwavesdotblog]

House Votes to Advance Bill That Could Ban TikTok in the U.S.

Legislation Passed: The House voted in favor of a bill that could lead to a ban on TikTok in the US unless ByteDance sells it to an American company. Senate Expectations: The bill, now heading to the Senate, is expected to pass there as well. Security Concerns: US politicians have security concerns over TikTok’s data sharing with the Chinese government, given ByteDance’s obligations. Potential…

[louistonehill]

A new tool lets artists add invisible changes to the pixels in their art before they upload it online so that if it’s scraped into an AI training set, it can cause the resulting model to break in chaotic and unpredictable ways. 

The tool, called Nightshade, is intended as a way to fight back against AI companies that use artists’ work to train their models without the creator’s permission. Using it to “poison” this training data could damage future iterations of image-generating AI models, such as DALL-E, Midjourney, and Stable Diffusion, by rendering some of their outputs useless—dogs become cats, cars become cows, and so forth. MIT Technology Review got an exclusive preview of the research, which has been submitted for peer review at computer security conference Usenix.   

AI companies such as OpenAI, Meta, Google, and Stability AI are facing a slew of lawsuits from artists who claim that their copyrighted material and personal information was scraped without consent or compensation. Ben Zhao, a professor at the University of Chicago, who led the team that created Nightshade, says the hope is that it will help tip the power balance back from AI companies towards artists, by creating a powerful deterrent against disrespecting artists’ copyright and intellectual property. Meta, Google, Stability AI, and OpenAI did not respond to MIT Technology Review’s request for comment on how they might respond. 

Zhao’s team also developed Glaze, a tool that allows artists to “mask” their own personal style to prevent it from being scraped by AI companies. It works in a similar way to Nightshade: by changing the pixels of images in subtle ways that are invisible to the human eye but manipulate machine-learning models to interpret the image as something different from what it actually shows. 

Continue reading article here

[jcmarchi]

The Sequence Chat #475: Ed Sim, Forbes Top Tech Investor, on AI Investing, Security, Agents and More

New Post has been published on https://thedigitalinsider.com/the-sequence-chat-475-ed-sim-forbes-top-tech-investor-on-ai-investing-security-agents-and-more/

The Sequence Chat #475: Ed Sim, Forbes Top Tech Investor, on AI Investing, Security, Agents and More

Founder of boldstart ventures and widely recognized one of the best early stage investors in the world, Ed shares his perspectives about the AI space.

The Sequence Chat is our series of interviews with top AI thought leaders and practitioners. We dive deep and don’t pull any punches 😉.

Today, we have another extra special interview. Ed Sim, the founder of boldstart ventures, is widely regarded as one of the best early-stage VCs in the world growing bodtart from $1M to over $800M. I’ve learned a lot from Ed over the years, and he has graciously agreed to share some of his thoughts about the AI space with us.”

You can subscribe to The Sequence below:

TheSequence is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

Let’s dive in:

Welcome to TheSequence. Could you start by telling us a bit about yourself? Share your background, current role, and how you got started in venture investing and AI.

Hi, my name is Ed Sim, and I’m the founder of boldstart ventures which we started in 2010. I’m on year 29 investing in technical founders reimagining the enterprise. I’ve seen and lived through a number of cycles from the Internet boom in the late 90s to the financial crisis in 2008 and now, with GenAI, the greatest platform shift that I believe I will ever see. Some of the founders I’ve been fortunate to partner with from Inception have started and built companies like Snyk (developer security), BigID (data security and privacy), Protect AI (AI security), Tessl (AI native software development platform), Superhuman, Kustomer (AI support), Blockdaemon, Front and many more.

I began investing in “AI” back in 2010, when the focus was on rules-based machine learning, and later rode the first wave of Robotic Process Automation (RPA) in 2017 (check out my post on RPA & F500). My venture investing journey started at JPMorgan, where I worked as an investment analyst building quantitative trading models using historical risk-based pricing data and Excel. While my role was essentially that of a data analyst, curiosity led my colleague and me to start recording macros to automate our tasks. This sparked a deeper interest, and we taught ourselves Visual Basic to push the boundaries of what we could achieve.

Within months, we were requesting the latest Pentium machines to handle larger computations, and I downloaded the Mosaic browser, diving into the early days of the internet. In 1996, I joined a VC fund in New York, marking the beginning of what has now been a super fun, 29-year adventure in venture investing and AI.

If you’re curious about what’s on my mind, I publish a weekly newsletter called What’s 🔥in IT/VC, where I share the latest trends and insights from venture capital, startups, AI, and security. I also delve into company building, raising capital, hiring, and navigating exits. As a huge fan of The Sequence, I’m excited to share my thoughts with all of you!

🛠 AI Work

You coined the term Inception Investing. Can you define what it means and explain how it aligns with the current dynamics of company building in generative AI?

Inception Investing is all about collaborating with founders before they even incorporate; helping them accelerate their ideation process, pre-selling the first hires, and leading that initial round of funding upon incorporation. This is not pre-seed or seed or any of that mumbo jumbo – it’s just straight out backing people who are highly technical who have a unique insight into what to build with a track record of having done so in the past. It can mean first time or third time founders. What’s unique is that an Inception round is unbounded by size. This, by the way, is super important because definitions of firms like pre-seed really imply smaller rounds and seed, slightly larger, and multistage, unbounded. However, founders just need one place to go to when they start a company, no matter the round size!

As you know, in the world of GenAI and because funds have too much money, an Inception round can be as big as $100M+ for experienced, in-demand founders. That being said, we meet founders who are thinking in two ways – either raise as little as possible and see where it goes like a CrewAI (initial round of only $2m) or raise a significantly larger amount like Tessl (initial round of $25M) for Guy Podjarny’s third company (Snyk valued at over $7B). When it comes to GenAI, you won’t see boldstart chasing those $100M rounds, but we have backed 2 stealth companies building specific foundational models with super experienced teams (Deepmind, Boston Dynamics…) and an ability to get/create proprietary data in robotics hand dexterity and bio research. Initial rounds were in the $10M range and the idea is to prove it before raising the mega >$50M next round.

The rapid growth of generative AI has redefined traditional fundraising trends, with seed rounds often reaching hundreds of millions of dollars. How has this shift impacted company-building and go-to-market strategies for early-stage startups?

Nail it, then scale it!

The amount of money you have should never change how you build your startup. When you receive your first dollar, the only goal is to build the best product possible which means having the right vision and initial engineering and product team to do so. Anything else is a distraction. No matter how much money a founder raises whether it’s $100k or $100M has to go through this same process – build a product and discover what product market fit is. You can’t spend your way to product market fit, nor can you skip steps.

For some the definition of what the minimum valuable product is can be different – some want to train their own model which is super expensive and can cost ten to hundreds of millions of dollars while others like a CrewAI can iterate with a small team before becoming one of the leading multi-agent frameworks out there, running >1M multiagent crews a day!

Either way, build a product, get to PMF, and do it as efficiently as possible. From first hand experience I have founders who have raised >$100M who still have a lean burn getting through the first stage and ready to ramp up spending when they get to PMF. Just because you have the money doesn’t mean you need to spend it; in fact, the more people you hire the slower you will go, so be super careful about ramping up too quickly.

Finally, because of all of the dollars flowing into AI-related startups and depending on the market they are going after, founders and investors do feel the bar to attract and pay for talent requires more capital. And while not thrilled about some of these massive rounds, I can concur that for some of these startups, there is no other way than to start with a war chest of dollars.

Regardless of your approach, my only advice is that too much money removes constraints, and the best founders are the most resourceful and creative when their backs are against the wall. If you have a large cash war chest, manufacture ways in your mind to make it feel like you don’t!

The initial wave of generative AI investments focused on “GPU-rich” companies that required billions of dollars upfront to experiment with their models. Many of these companies struggled to gain meaningful traction, leading to pivots or acquisitions. How has this affected venture capital perspectives on the AI space? Are we moving away from the GPU-centric approach?

Well, those opportunities sucked up lots of money and many did not end well. Goes back to my point above about too much cash. We are way beyond the “let’s build our own general purpose” model phase now. We know who the leaders are in the general purpose LLM game and investors and founders are seeing that value is accruing up the stack where companies are much more capital efficient and can deliver value to the end user. It’s portfolio companies like Clay or Superhuman which are using OpenAI and Anthropic but then building their own twist for outbound data enrichment or email to grow insanely fast. It’s companies like Anysphere, creators of Cursor, and other AI-native software companies who are growing rapidly that are attracting the next big dollars in venture.

Finally, I still believe that the last mile in the enterprise is the longest mile. There is so much to get right besides choosing what model to use – how do you make sure only the right data can be seen by the right user, how do you make sure the right prompts are used to get the best answer, how do you remove hallucinations, how do you deliver on-premise…you get the idea. Investors and founders are also building vertically focused-AI companies whether in finance, law or HR – no industry is immune as this GenAI wave is bigger than just SaaS.

If you believe that software ate the world and AI is eating software then you have to ask the question if GenAI will eat into labor. Because if you believe that last point, the opportunity to transform labor markets and capture some of those dollars is in the multi-trillions – this is the opportunity we are all chasing.

You’ve been a successful investor in enterprise AI security and have emphasized that “there is no AI without AI security.” How do traditional cybersecurity practices and techniques need to evolve in the era of generative AI?

We can hit this from lots of different ways. First, hackers are always one of the first adopters of technology and of course, they are doubling down on AI. It’s always easier to attack than it is to defend. Cybersecurity practitioners need to double down on threats generated by AI which is especially great at social engineering like emails, fake voice and video calls. In addition, GenAI allows hackers to easily send these messages at scale and also do other things like probe networks, find new software vulnerabilities, and even find new ways to hack into systems.

Secondly, we need to think about second order effects using AI. The more code that is written by AI the more one needs to analyze to secure that code. Companies like Snyk in our portfolio scan and secure code as AI writes it. The basics like LLM prompt injection attacks are getting taken care of, but the next wave is agents – who’s going to make sure agents are acting against the right policies and who’s going to provide the infrastructure for agents to authenticate and validate who they are.

Finally, any AI models being used in an enterprise and embedded in applications open up opportunities for hackers to exploit. If you think of the SBOM or software bill of materials then we have the AIBOM which is the AI Bill of Materials which is even more complex as it not only includes software but also data and the model itself. For example, one of our portfolio cos where I’m on the board, ProtectAI, has a partnership with Hugging Face to scan all of the open source models for vulnerabilities.

At the end of the day, cybersecurity professionals need to look at the threats from GenAI holistically from the network to the software to the data to the people and ultimately use AI to keep up with all of the AI threats coming at scale! Many of the smartest CISOs I know are already tinkering with agents to automate some of their response workflow, and I expect this to become more prevalent in the next couple of years so these cybersecurity teams can continue to scale while hiring less.

Enterprise security is known for being a challenging market with high costs, long sales cycles, and significant resistance to change. Is AI security innovative enough for new startups to disrupt the market, or will the incumbents continue to dominate?

Many of our best Inception investments in cybersecurity were made in founders who could see the future and anticipate new attack vectors that hackers could exploit. Despite how long and hard it can be to sell into these organizations, every CISO has some discretionary budget to spend on new threats. In my 29 years as an enterprise VC, I’ve never seen a category, if you will, explode in interest as fast as AI security. One caveat is that the idea of AI security is super broad and for these purposes I want to limit it to securing AI usage in the enterprise. This covers the SaaS apps that employees use, the models and software that enterprises build and deploy, and of course the data security and privacy around it.

As I lay this out you can already see the breadth and depth needed to be an AI security vendor, and I can promise you that no incumbent vendor other than Microsoft has the understanding to cover every category. Because of that, you are seeing niche startup vendors play in LLM and prompt injection security, network security, AI model security from offensive red teaming to understanding the AIBOM as mentioned above, data privacy, open source security, agent security… you get the picture. Sure some incumbents like Cisco bought Robust Intelligence and actually released something interesting called Cisco AI Defense. And of course we have Protect AI, a boldstart portfolio company, which is the only pure play startup that covers the full gamut of AI security from what developers build to what AI engineers use for models to what employees use and finally the AI-SPM dashboard for CISOs. In my opinion, this category will rapidly consolidate as niche vendors either get swallowed up or killed by incumbents who have a large installed base of customers while a select few startup vendors will reach escape velocity as one of the standalone new platform plays for AI Security.

Boldstart has been active in the AI agent space with investments like Crew AI. What are your views on AI agents, and how can companies succeed in such a hyper-fragmented market?

Yes, we’ve been fortunate to have a front row seat in the agent space as we led CrewAI’s Inception round and have watched the team scale to over 1 million multiagent crews run per day. The growth has been simply astounding. More importantly, CrewAI is not just a platform to build multi-agent teams but also to build and orchestrate agentic workflows needed to solve business problems. This requires offering a developer an agent system that autonomously performs tasks with minimal human intervention by observing, planning, acting, learning and repeating. CrewAI helps developers easily build these teams and workflows, and I feel like we are just in the first inning of a long game.

There are tons of competitors appearing every single day from established vendors like Microsoft and Google to a number of new startups. I don’t know how this will all shake out but what I do know is that winning requires a founder with a killer long-term vision, an amazing product which developers love as winning the hearts ❤️ and minds 🧠 is what’s required for success, speed of execution, and finally building out an ecosystem of partners who use your framework as one of their core offerings. Fortunately CrewAI hits the bullseye 🎯on all fronts and was just announced as a key Nvidia partner at CES by Jensen Huang as he laid out his vision of the agentic future. Other partners include Cloudera and IBM with more to come.

In the long run, I also believe we will live in a world where each of us will have hundreds of agents doing work for us around the clock and if you extrapolate the second order effects, we can easily conclude that a whole new infrastructure will be needed to support hundreds of thousands of agents doing work at enterprises. Think about runtime ephemeral authentication and access, observability for all of these heterogeneous agents, interoperability for these agents to talk to one another, security policies on what is allowed and not, and governance. These are all areas we at boldstart are investing in or have already invested in 😄with companies in stealth.

In the AI agent space, which segment is better positioned for short- to medium-term success: vertical, domain-specific agents or horizontal, agentic platforms?

In the long run we will all win if agents get smarter, hallucinate less, and are easily programmable and monitored. We have to remember that this enterprise AI wave is MUCH BIGGER THAN JUST SOFTWARE. AI is Eating Software but will AI also EAT LABOR? We are not just talking about reallocating existing SaaS dollars. AI’s impact on worker productivity could be an additional $6-8 Trillion, and if vendors can capture some of that value that is enormous potential.

When it comes to who wins over time, there will be winners in every category from vertical to domain to agentic platforms. We’re investing in all of these areas from the infrastructure to more domain experts building armies of agents to help security professionals detect and respond faster to security threats to automating customer service to … In the end, it’s super easy to start a company now but hard to build multi-hundred million revenue businesses. For every 500 companies that get funded, maybe 1-2 reach escape velocity. The amount of companies that don’t make it will be higher than ever, but for the companies that do survive, they will create far more value than all of the money lost. That’s the opportunity!

boldstart has had remarkable success in the SaaS space. Do you think traditional SaaS will be replaced by AI-driven agentic experiences? Are we transitioning from form-based interfaces to multimodal agent experiences?

Once again I don’t think this is an all or nothing proposition. It’s pretty clear that if agents proliferate like we expect them to then it will reduce the amount of seats in an enterprise and will cannibalize seat-based pricing which is the de facto business model for existing SaaS applications. Depending on the function, we will eventually have outcome-based pricing, no doubt. In fact, one of our boldstart portfolio companies, Kustomer, was one of the first customers to offer 100% outcome based pricing which you can read about here. When we did the math, it ate into short-term revenue marginally while future proofing us for the long term. The customers frankly love it.

Satya Nadella from Microsoft was recently interviewed and said that “the notion that business applications exist, that’s will probably all collapse in the agent area. They are essentially CRUD databases w/biz logic. All logic going to Agents which are multi-repo, multi-vendor. When logic moves to an agent, then people will replace back-ends. Seeing high win rates for Dynamics with agents.”

I agree with him 100% but my question is how long will this take?

Besides agents accessing applications or just the raw data, I 100% believe we will have other interfaces like voice take off or even using your camera to help field workers get work done. Chat based interfaces are nice, but we will see that this will eventually become a relic in the years to come as machines talk to more machines and we have other ways of accessing data for systems. I’m also interested in the idea of dynamic UIs for each user or role or based on the context or location where the application and what you see may change every single login, intuiting what you want to see and just giving you the answer.

The next few years is going to be interesting to watch MSFT, Salesforce, ServiceNow and others protect their turf, offering their own agentic workflows, while startups build from a clean slate with no app or dog 🐶 in the hunt. Regardless, the opportunity to reshuffle the decks in the years to come is mind blowing! Hundreds of billions of dollars of SaaS revenue are at stake and then there will be trillions of dollars of opportunity if these agents eat into the labor market!

💥 Miscellaneous – a set of rapid-fire questions

What’s your favorite area of generative AI outside of security and agents?

I feel like a kid in a candy shop at the moment. I’m just enjoying playing around with all of the new personal tech coming around and at the moment Google Notebook LM is giving me insane super powers allowing me to easily research and understand research papers, documents, and synthesize podcasts and turn all of that into a podcast with Q&A – just so much fun.

Do you think we’ll achieve AGI through transformers and scaling laws, or will it require entirely new architectures?

That’s above my pay grade 🤣, and I’ll let the scientists decide. But at the moment, it seems to me we still have some room to grow using transformers by leveraging agentic reasoning and scaling test time compute. We also have other new promising areas of research like Google Deepmind’s Titan for long-term memory to deliver even better answers faster. That being said, all technology eventually gets displaced!

What advice would you give to founders starting in the generative AI space? What’s the most common mistake you’ve seen founders make in AI companies?

Think in first principles – focus on the problem you are solving and for whom, how are you uniquely solving that problem to make an end user’s life orders of magnitude better with your product or service than without, and then think about the AI last. If you think about AI first, you can get lost in the jungle focusing on a cool technology looking for a problem to solve. Then think about your data moat or the long term secret sauce of your business if successful and as you scale – too many folks are rushing to get an AI product out the door which can easily be copied with no long term defensibility. Finally, remember that the “perfect is the enemy of the good” which means you should ship product and iterate as fast as possible. Also think about how you use AI internally for development, sales, customer support, and outbound to build as lean a company as possible with as little venture money as possible 😄.

Who is your favorite mathematician or computer scientist, and why?

Hands down Claude Shannon as without him we wouldn’t have Information Theory which is the concept of encoding and transmitting information efficiently. This theory underpins how data is processed, stored, and transmitted, which is critical for pretty much all technology today, especially AI systems. In addition he was one of the first to think about building machines that could think, developing an early chess playing program which showcased how machines could make decisions. Finally we wouldn’t have the freedom we have today with wireless communications without Claude Shannon!

[probablyasocialecologist]

Artificial intelligence is worse than humans in every way at summarising documents and might actually create additional work for people, a government trial of the technology has found. Amazon conducted the test earlier this year for Australia’s corporate regulator the Securities and Investments Commission (ASIC) using submissions made to an inquiry. The outcome of the trial was revealed in an answer to a questions on notice at the Senate select committee on adopting artificial intelligence. The test involved testing generative AI models before selecting one to ingest five submissions from a parliamentary inquiry into audit and consultancy firms. The most promising model, Meta’s open source model Llama2-70B, was prompted to summarise the submissions with a focus on ASIC mentions, recommendations, references to more regulation, and to include the page references and context. Ten ASIC staff, of varying levels of seniority, were also given the same task with similar prompts. Then, a group of reviewers blindly assessed the summaries produced by both humans and AI for coherency, length, ASIC references, regulation references and for identifying recommendations. They were unaware that this exercise involved AI at all. These reviewers overwhelmingly found that the human summaries beat out their AI competitors on every criteria and on every submission, scoring an 81% on an internal rubric compared with the machine’s 47%.  Human summaries ran up the score by significantly outperforming on identifying references to ASIC documents in the long document, a type of task that the report notes is a “notoriously hard task” for this type of AI. But humans still beat the technology across the board. Reviewers told the report’s authors that AI summaries often missed emphasis, nuance and context; included incorrect information or missed relevant information; and sometimes focused on auxiliary points or introduced irrelevant information. Three of the five reviewers said they guessed that they were reviewing AI content. The reviewers’ overall feedback was that they felt AI summaries may be counterproductive and create further work because of the need to fact-check and refer to original submissions which communicated the message better and more concisely. 

3 September 2024