Spencer Thomason

Startup Hakk

About the show

Welcome to StartupHakk. My name is Spencer Thomason. I am the CEO and Co-founder of CleanRouter, as well as many other products. Here are some lessons that I've learned building a successful business while challenging startup norms. My challenge is to push you to rethink startup success.

Newest episodes

DeepSeek Is Giving Away Its Entire AI Stack!

Oct 22, 2026 9 min

DeepSeek reportedly reached a $1 billion annualized revenue run rate while giving away its AI models and infrastructure. Meanwhile, OpenAI is pursuing expensive subscriptions and enormous infrastructure spending.The numbers are interesting. DeepSeek reportedly achieved an 82% API gross margin, compared with OpenAI's reported 39% gross margin in the first quarter.But the bigger story is the strategy.DeepSeek is releasing model weights, research, and even its inference framework, betting that it can win by offering efficient, low-cost AI rather than locking everything behind subscriptions.In this video, I break down why I think DeepSeek is playing a different game from OpenAI and Anthropic, what its potential Huawei chip partnership could mean, and why we believe in the same open-source, local-first approach with OpenMonoAgent.Can DeepSeek's open-source strategy beat OpenAI's subscription-based business model?

OpenAI and Anthropic Have a Trillion-Dollar Problem

Oct 20, 2026 9 min

The AI industry could need $10 trillion in infrastructure investment over the next decade. Amazon says demand for AI computing is stretching years into the future, but there's a much bigger question: who's going to pay for all of this?OpenAI and Anthropic are growing fast, but the cost of building and operating this infrastructure is staggering. Some estimates suggest data centers would need to generate $8.5 trillion in revenue to justify the investment. I'm not convinced the financial math works.I think we're eventually going to hit a funding wall. That doesn't mean AI disappears. The technology is here to stay, and open-weight models could give smaller businesses a major advantage without depending on expensive subscriptions and cloud providers.So what do you think? Is this the biggest financial bubble in tech history, or are we underestimating how much AI demand will grow?

Anthropic's Financial Numbers Are A HUGE Mess!

Oct 15, 2026 9 min

Anthropic's revenue is growing fast. But I think there's a much bigger problem hiding behind those numbers.We're talking about a company that reportedly has more than $518B in future AI compute commitments.That's an enormous amount of money to spend on infrastructure, especially when the revenue growth may already be slowing.And here's what really bothers me.Anthropic keeps talking about being profitable on an adjusted basis.Adjusted profits. Adjusted margins. Adjusted cash flow.Okay, but what does the business actually cost to run?You can't keep celebrating revenue growth while ignoring the enormous cost of building and operating these AI models.Then there's Amazon and Google.They're investors, infrastructure providers, and major distribution partners for Anthropic. That's a lot of dependence on companies that are also building competing AI products.And now we're hearing about potential IPO valuations approaching $2T.I don't buy it.I think Anthropic is going to have a very difficult time getting through the financial scrutiny that comes with going public. And even if it does, I don't see how a $2T valuation makes sense based on the numbers being discussed.I've said it before. Revenue is not profit, and growth doesn't automatically mean you've built a sustainable business.The AI industry needs to start paying attention to what these companies actually spend, not just what they bring in.Would you invest in Anthropic at a $2T valuation, or do you think the AI bubble is getting out of control?

Similar podcasts

Founders David Senra

Raw Talks With Vamshi Kurapati - Telugu Business Podcast Vamshi Kurapati

WTF is with Nikhil Kamath Nikhil Kamath

The Tim Ferriss Show Tim Ferriss: Bestselling Author, Human Guinea Pig

My First Million Hubspot Media

The Prof G Pod with Scott Galloway Vox Media Podcast Network

Good Bad Billionaire BBC World Service

声动早咖啡声动活泼

All episodes

The newest 50 of 314 episodes

DeepSeek Is Giving Away Its Entire AI Stack!

Oct 22, 2026 9 min

OpenAI and Anthropic Have a Trillion-Dollar Problem

Oct 20, 2026 9 min

Anthropic's Financial Numbers Are A HUGE Mess!

Oct 15, 2026 9 min

Anthropic Just Cut Off Its Own Paying Customers

Oct 14, 2026 10 min

Anthropic just made a major move to stop other AI labs from copying Claude, but there's a catch that could hurt its own business. The company says its new defenses have reduced the amount of Claude's reasoning that competing labs can extract by up to 99%. Tens of thousands of accounts have been blocked, and trillions of tokens of suspected distillation traffic have been turned away. Here's where things get interesting. Those accounts weren't necessarily using Claude for free. Some were paying customers generating enormous amounts of API traffic. So what happens to Anthropic's revenue when that activity disappears? The company is trying to protect its intellectual property, which makes sense. But cutting off paying traffic could affect growth expectations, especially with questions surrounding a potential IPO. And there's another problem. Paying Claude customers are also losing access to some reasoning details, while certain blocked requests can still incur charges. At StartupHakk, we've been advocating for local-first AI because businesses shouldn't have to depend entirely on models and pricing rules they don't control. Anthropic may be protecting Claude, but the financial consequences could be much bigger than people expect. Do you think Anthropic is making the right move, or could this crackdown end up hurting its business?

Microsoft and Nvidia Are Building AI Beyond OpenAI

Oct 13, 2026 9 min

Microsoft is developing its own 137B-parameter coding model and bringing powerful AI agents directly to Windows PCs. With Nvidia's new hardware and GitHub Copilot's hybrid approach, more AI workloads could run locally instead of relying entirely on the cloud.Here's what's interesting.Microsoft spent years helping OpenAI grow. Now it's building its own models and infrastructure that could reduce its dependence on OpenAI. Meanwhile, Nvidia wins on both sides by supplying the hardware for cloud AI and local machines.But there's a catch. Running large AI models locally requires serious memory, the most powerful hardware isn't cheap, and Microsoft's approach still depends on the cloud for certain tasks.At StartupHakk, we've been advocating for local-first AI because businesses should have more control over their models, data, and infrastructure.The question is whether Microsoft can make local AI powerful and affordable enough to change how we use AI.Would you buy a Windows PC specifically to run AI locally, or would you rather stick with cloud AI?

Microsoft CEO Reveals AI's Biggest Security Problem

Oct 13, 2026 11 min

Microsoft CEO Satya Nadella just made a serious point about AI security: businesses should treat AI models like potential insider threats.And Microsoft is already acting on it with new Execution Containers designed to keep AI agents isolated.OpenAI Codex is also adopting stronger sandboxing controls.I've been saying this for a long time. AI agents shouldn't have unrestricted access to your files, code, credentials, and business systems.At StartupHakk, we built OpenMonoAgent around sandboxing, command logging, and local-first AI because security shouldn't be something you add after something goes wrong.The future of AI isn't just about smarter models. It's about controlling what those models can actually do.Would you trust an AI agent with unrestricted access to your company's data?

Microsoft's 6,000 H-1B Filings Under Fire

Oct 9, 2026 8 min

The White House just accused Microsoft of abusing the H-1B visa system while American tech workers face layoffs. Now Microsoft, Adobe, and six major outsourcing firms have been suspended from the PERM green card program. But here's the twist. Hours after the crackdown, President Trump awarded Microsoft CEO Satya Nadella a national technology medal. Microsoft says 80% of its roughly 6,000 H-1B filings were for existing employees, not new hires. I've heard complaints about this system for years. American workers deserve a fair shot at these jobs, but immigrant employees shouldn't be trapped in a system where their future depends on one employer. And despite the headlines, this isn't a blanket H-1B ban. It's a freeze on a critical part of the green card process. So is the government actually fixing the problem, or are workers on both sides paying the price? What do you think? https://StartupHakk.com

Meta Is Betting Big on Muse. And It’s Winning.

Oct 8, 2026 13 min

Meta Is Betting Big on Muse. And It’s Winning. $27B a year. That's what Citi reportedly thinks Meta Muse could generate by 2030. Deutsche Bank's upside case goes even higher. And Meta doesn't have to play the same game as OpenAI and Anthropic. Instead of making money every time you burn another pile of tokens, Meta can give away enormous amounts of AI usage and take a small cut when Muse helps you actually buy something. Meta already has the users. It already has the distribution. It already has the advertising machine. Now Zuckerberg may have found the AI product to put on top of it. I've bashed plenty of his decisions in the past, but I'll give him this one: Muse is a real product. Would I use it? Absolutely. Would I give Meta my production data, banking information, or sensitive company data? No. For that, I still want AI running on infrastructure I own and control.

Microsoft and Meta Just Cut Claude Use Before Anthropic’s IPO

Oct 6, 2026 10 min

Microsoft had reportedly been on track to spend more than $1B a year with Anthropic. Some engineers inside Microsoft's Cloud and AI group reportedly had AI budgets reaching $100,000 a month. Now Microsoft is cutting those budgets, reducing Claude spending, and directing employees toward GitHub Copilot and its own ecosystem. Meta is doing something similar. Tens of thousands of Meta employees had been using Claude Code, but Meta has been expanding its own internal AI coding tools instead. And all of this is happening as Anthropic moves toward a potential IPO with 2 customers reportedly accounting for roughly 25% of its revenue. That's worth paying attention to. Because while some of the biggest companies in technology are learning that renting massive amounts of AI compute gets expensive very quickly, Anthropic is simultaneously warning about the dangers of powerful open-weight models. We see the open-weight story differently. Yes, powerful models can be misused. But open weights also let defenders run models against their own code, keep sensitive data inside their own infrastructure, modify the system, and build tools without depending on a closed provider to decide what they're allowed to do. Microsoft and Meta can afford to build their own AI infrastructure. The rest of us should at least learn how to own more of ours. Stop thinking of AI as something you have to rent forever. Learn the stack. Own the infrastructure.

We Are So Early in AI

Oct 5, 2026 9 min

91% of households pay for streaming.55% pay for cloud storage.Only about 2% pay for AI.For all the talk about AI taking over everything, almost nobody is actually writing a monthly check for it yet.Meanwhile, we're pouring enormous amounts of money into GPUs, semiconductors, data centers, and AI infrastructure.That's not a victory lap. It's a bubble check.The AI boom feels enormous if you live inside it every day, but the numbers show just how early we really are. Mass adoption hasn't happened yet—and when it does, it may look less like everyone buying a chatbot and more like AI disappearing into the software we already use.If you feel behind, you're not. Most people haven't even started.

OpenAI Was Warned. Then It Fired Them.

Oct 3, 2026 10 min

OpenAI employees reportedly warned leadership that its newest AI models weren't being monitored closely enough during testing. Then OpenAI parted ways with 3 people for allegedly mishandling sensitive company information outside established procedures. OpenAI says policies were violated and trust was broken. But I think the bigger question is what caused people inside the company to raise those security concerns in the first place. These systems don't magically “go rogue.” Humans give them compute, tools, access, permissions, and objectives. If the monitoring isn't strong enough, that's a security problem. And as AI systems become more powerful, security can't be something we deal with after deployment. Scan your systems. Review your code. Find the vulnerabilities before somebody else does.

The Most Dangerous Thing About AI Isn’t What You Think

Oct 2, 2026 15 min

27% → 9% accuracy. 30% → 76% confidence. 44% → 3% willing to say, “I don't know.” That's the AI problem I'm much more concerned about than some science-fiction rogue agent. AI can be completely wrong while sounding completely certain—and the person using it can walk away more confident because of it. These systems are designed to be agreeable. Give them an idea and they can validate it, reinforce it, and keep escalating instead of telling you that you're wrong. That's why I never ask AI to tell me where I'm right. I ask it to find where I'm wrong. Find the holes. Find what I missed. Challenge the assumption. I'm not anti-AI. We build with it every day. I'm an AI realist. AI isn't going away, so we need to understand how it works, keep humans in the loop, and learn to use it without outsourcing our judgment to it.

Anthopic's IPO - Not Profitable. Never Profitable. Compute Insane

Sep 30, 2026 9 min

Anthropic’s IPO Math Is Insane$4.6B in revenue.More than $8B in operating losses.Nearly $7B spent on compute.And potentially hundreds of billions more in cloud compute through 2030.Yes, Anthropic is growing fast. And the nearly $42B headline net loss needs context because roughly $34B came from an accounting charge.But look at the underlying business.Compute requirements keep exploding while the improvements between new models keep getting smaller. We're pouring dramatically more resources into these systems without getting dramatically better models back.That's diminishing returns.Now Anthropic wants the public markets to buy into these economics at an enormous valuation. I think this AI bubble is hugely inflated, and I don't think this math works forever.That's one of the reasons we built OpenMonoAgent. I want AI infrastructure we can actually own and control instead of permanently renting somebody else's compute, API, and pricing model.

Jacob Coxon Lied — “AI Fear” Is Part of a Coordinated Psyop

Sep 30, 2026 12 min

18 minutes. That's how long after a Wall Street Journal exclusive quoting Jacob Coxon went live that his own resignation thread appeared. Coxon presented himself as one researcher walking away from Anthropic and sounding the alarm about AI. Then reporting surfaced around PR bookings, overlapping interview calendars, and other AI safety figures being lined up around the same rollout. Our argument: this wasn't simply one guy having a midnight crisis. It looked like a coordinated media launch. And the bigger issue is what happens next. If fear becomes the justification for deciding who is allowed to run powerful AI, compute and inference could become concentrated in the hands of a tiny number of companies. We think the answer goes the opposite direction: keep AI open, let people control their own infrastructure, and build security around systems people actually own. What do you think—genuine grassroots AI alarm, or coordinated campaign?

Sam and Dario Belong in Jail — “Rogue” Agents Don’t Run Themselves

Sep 29, 2026 17 min

Tens of thousands of AI incidents are now reportedly being investigated across OpenAI, Anthropic, and security researchers.And somehow we keep hearing the same phrase: “rogue agents.”But an AI agent doesn't give itself millions of dollars worth of compute. It doesn't give itself internet access. It doesn't design its own sandbox. It doesn't decide what tools it gets or what systems researchers point it toward.People do.That's the part getting lost in the panic.If you build the system, give it extraordinary resources, point it at the open internet, and fail to properly contain it, you don't get to act surprised when something goes wrong.Our argument isn't that AI security doesn't matter. It's the exact opposite.Stop blaming the car. Start holding the driver responsible.And if you're running AI against real systems, build the security and containment correctly from the beginning.

Microsoft thwarts massive cybercrime ring! Protect yourself today!

Sep 25, 2026 11 min

12,000+ compromised inboxes. More than 10,000 organizations. AI digging through stolen email, mapping trusted relationships, figuring out who controls the money, and helping criminals build impersonation fraud.Microsoft just helped disrupt EvilTokens, and this is a pretty clear look at where cybercrime is heading.Once attackers got access to an inbox, AI could help do the homework: understand the organization, identify valuable relationships, and turn that information into a roadmap for fraud. Microsoft thwarts massive cyber…But Microsoft is also using AI on the other side.Instead of throwing the biggest model at every security problem, they're using specialized models for most of the work and escalating the hardest cases.Model. Data. Harness.That's the important part.Attackers are already figuring out how to build systems around AI. We're following the same model-data-harness pattern on defense with StartupHakk Security, OpenMonoAgent, and our security playbooks.Security isn't a magic checkpoint. Security is the harness. Run a free site scan or code review by visiting: security.startuphakk.com

OpenAI and Anthropic will NEVER IPO

Sep 24, 2026 10 min

Alex Karp questioned whether the liability surrounding frontier AI could make that path far more complicated.Both companies have already submitted confidential draft S-1s, so the process has started—but that doesn't guarantee either ultimately completes an IPO. Our prediction is that they don't. At the same time, open-weight models keep getting better, and the model itself is becoming only part of the equation. The harness and infrastructure around it matter. That's why we built OpenMonoAgent. AI shouldn't be another subscription you rent. It should be infrastructure you own.

Claude Is Getting Dumber — Anthropic Is Lying to You!

Sep 23, 2026 10 min

Claude works great one day. The next day it feels noticeably worse.People have been complaining about it for a while. Now somebody actually measured what was happening.Even with high or max effort selected, the reported amount of thinking delivered across calls varied dramatically. Some calls received little thinking. Others reportedly received none.And longer reasoning runs rarely reached the levels associated with the published benchmarks.That's the real problem with renting your AI.You pay for the model name.But somebody else controls what happens behind that name.That's why we built OpenMonoAgent around a completely different idea: own the stack.Your hardware.Your model.Your data.Your inference.For Free.

Zuck’s Meta Muse Got Hacked!

Sep 22, 2026 9 min

Meta spent weeks talking about the privacy and security behind Muse. Then a zero-day dropped showing how a local process with no special privileges could hijack Muse on Mac. And this isn't just another chatbot. Muse can be given access to files, mail, messages, calendar, microphone, camera, location and other connected services. The disclosed vulnerability could redirect dictated prompts to an attacker-controlled server and potentially expose authentication material. That's exactly why giving AI agents enormous privileges without properly containing them is dangerous. Privileges without a locked harness aren't a feature. They're a gift to whoever gets code running on the box first. We prefer AI infrastructure where the agent is contained and its privileges are controlled. That's how we've approached OpenMonoAgent, and it's also what powers our StartupHakk Security tools. Want to see what vulnerabilities might already be sitting in your website or code?

They Hacked OpenAI With Claude - Easy Way to Check Your Security!

Sep 21, 2026 10 min

OpenAI got hacked. And the security researchers behind it used Claude to help them do it. But this wasn't some rogue AI that suddenly decided to attack OpenAI. A skilled security team picked the target, used Claude to accelerate the difficult parts, chained 2 vulnerabilities together, gained access to an OpenAI employee's ChatGPT and Codex accounts, and proved the access with a harmless pull request inside an internal repository. Humans steered it. AI accelerated the hard parts. That's what matters. Security researchers are starting to push increasingly capable AI models to see what they can do. For skilled teams, these models are becoming incredibly powerful tools. So taking those tools away from defenders isn't the answer. We need to make sure the good guys have the tools they need to defend against the bad guys. That's why we built StartupHakk Security. You can scan your website or upload your codebase and start looking for vulnerabilities immediately. Get your top 3 findings free, with deeper security reviews and penetration testing available when you need them.

StartupHakk Security - Free Scans for the AI Age

Sep 18, 2026 5 min

How many vulnerabilities are sitting inside your website or codebase right now that nobody knows about?The average data breach now costs companies almost $5M, and the number in the U.S. is around $11M according to the figures discussed in this demo.But finding security problems shouldn't require waiting until after something goes wrong.We're walking through StartupHakk Security from beginning to end: create an account, launch a website or code scan, watch it run, review the findings, see the evidence, work through remediation, export the results, and re-scan when you're ready.The goal is simple: know what's exposed before someone else finds it.If you want to protect your business from vulnerabilities hiding in your website, codebase, or technical infrastructure, check out StartupHakk Security

AI Models “Escaped” 5 Labs in 30 Days - Here’s What Actually Happened

Sep 17, 2026 11 min

5 AI labs reported models “escaping” containment in 30 days. One agent ran for 4.5 days and executed roughly 17,000 actions. Anthropic ran around 141,000 evaluations and found cases where models reached real production systems.That sounds terrifying.It also doesn't tell the whole story.A lot of these incidents trace back to the same third-party testing environment. The models were told the environment was sealed and they had no internet access.It wasn't.There were open paths, and the agents found them.In several cases, these models didn't invent some extraordinary new exploit or suddenly develop a desire to “escape.” They were given a goal, found access that shouldn't have been there, and used it.There's also something worth noticing about where these stories are coming from.The dramatic announcements have largely come from the AI labs themselves rather than independent whistleblowers. That doesn't make the incidents fake, but it does mean we should separate the actual technical evidence from the “our AI is becoming too powerful to control” narrative surrounding it. That's the skepticism we're bringing to these reports.But here's the part that isn't funny:Companies are starting to give these agents real access.Codebases. Databases. CI/CD pipelines. Production clusters. Internal systems.An AI agent doesn't need to “go rogue” to become dangerous.Give it a clear objective, too much access and one security hole, and it may simply use what you gave it.That's the real lesson from these AI escape stories.And if you want to protect your business from exposed vulnerabilities, insecure code and the attack paths that increasingly capable AI agents can exploit, check out StartupHakk Security.

Zuck says - Dario and Sam: Step Aside If You Cannot Control It!

Sep 16, 2026 12 min

If you genuinely believe the AI you're building could cause catastrophic harm, why are you still shipping it? That's the contradiction underneath the latest AI safety fight. Dario Amodei is warning about frontier AI. David Sacks' response is much simpler: if you can't control what you're building, slow yourself down or step aside. Don't make everyone else stop because you don't trust your own product. We're much more interested in the builder approach: own the risk, build the guardrails, and keep humans in control. Microsoft AI's newly published Code of Conduct takes an explicit human-control approach, including requirements around interruption, correction and shutdown. Meta has likewise described safety testing and deployment controls for its advanced models. That's also why we care so much about local AI. OpenMonoAgent.ai is open source, runs on your hardware and gives you control of the stack instead of putting everything behind somebody else's API. AI shouldn't be something you fear while continuing to ship it. Build what you can control.

OpenAI and Anthropic Valuations Are RIDICULOUS!

Sep 15, 2026 14 min

OpenAI and Anthropic are reaching valuations that are getting harder to take seriously. The problem isn't whether their technology is impressive. It obviously is. The question is what you're actually valuing when these companies are still burning enormous amounts of cash while companies like Meta already have profitable businesses capable of funding massive AI development internally. Spencer thinks people are seriously underestimating Meta here. We're breaking down the numbers, the IPO problem, and why the economics behind frontier AI may look very different once investors can see everything. And as open models keep improving, there's another question worth asking: how much of your AI stack do you actually need to rent? Watch the full breakdown.

Dario is crazy - why legislation won't fix anything

Sep 14, 2026 16 min

Dario Amodei says we need to “pace the frontier.”Sam Altman agrees.Then Elon Musk says “Dario is right.”Three people with very different positions on frontier AI suddenly landed on essentially the same message over one weekend.The question isn't whether AI has risks. Of course it does.The question is who gets to write the rules.If AI regulation becomes expensive and complicated enough that only the biggest frontier companies can satisfy it, we haven't made AI safer. We've made it harder for startups, open-weight models and independent builders to compete.That's why local and open AI matter.With OpenMonoAgent.ai, we're building around the opposite idea: run models on infrastructure you control, keep ownership of your data and code, and stop treating access to intelligence like something that has to be rented forever.Let people build. Let the market compete.

Anthropic should be ashamed! Coxon to Capitol in 48hours

Sep 11, 2026 13 min

An almost unknown AI researcher quits Anthropic. Minutes after a Wall Street Journal exclusive, he posts that AI could kill us all within years—and the post explodes to more than 140 million views despite almost no previous posting history. Then the story gets picked up by AI-safety groups, media, and politicians. We're looking at why this sudden wave of AI fear deserves scrutiny, and what happens when existential-risk arguments start turning into calls for regulation that could make it harder for open-weight and locally controlled AI to compete. Because if the rules become expensive enough that only the biggest frontier labs can comply, we haven't decentralized AI—we've concentrated it. That's exactly why we're building OpenMonoAgent.ai. Run AI on your own hardware, keep control of your code and data, and stop depending entirely on somebody else's models, infrastructure, and pricing. AI shouldn't be something a handful of companies get to control.

OpenAI's Trojan horse - more ChatGPT stealing coming to light!

Sep 10, 2026 11 min

OpenAI is giving 10,000 top scientists, mathematicians, and engineers free access to its frontier models, with plans to expand toward 100,000 researchers through 2027. It sounds incredibly generous.But what happens when the thing being typed into those models isn't a normal ChatGPT prompt—it's unpublished research, original mathematics, proprietary reasoning, or intellectual property?That's the question we're digging into.New allegations from mathematicians are raising questions about whether unpublished work shared through AI conversations may have contributed to mathematical breakthroughs later attributed to OpenAI's models. We're also looking at the reported German website incident involving OpenAI agents and the broader pattern these stories raise around data, ownership, and control.The bigger lesson isn't just about OpenAI.If your most valuable work is being processed on someone else's infrastructure, you're trusting someone else with it.That's exactly why we've been pushing local AI with OpenMonoAgent.ai: run the models on hardware you control, keep your code and data local, and own the infrastructure instead of renting access to it.Watch the full breakdown and tell us: would you put unpublished work into a frontier AI model?

Why Claude is NOT the future of Legal AI

Sep 9, 2026 7 min

Claude usage among legal professionals more than doubled in a year — from 15% to 33%.That's a huge shift.But here's the part that matters: 57% cite data privacy and confidentiality as a barrier to AI adoption, while 46% cite hallucinations.Lawyers want AI. The question is whether another general-purpose chatbot is actually the answer.Open a laptop at a small firm and look at the workflow.Case management in one tab.Claude in another.Email somewhere else.Documents somewhere else.Signatures somewhere else.Now the lawyer is the integration layer, copying information between systems and putting sensitive client data into tools that weren't built around the matter.That's the problem.Claude can write. But it doesn't run your law firm.SwiftCase Legal AI is built around the matter from the beginning—case files, AI-assisted analysis with citations, clients, signatures, time tracking, billing, and the rest of the workflow in one system.The future of legal AI isn't another chat window.It's software that actually understands the work.

Skills Are a Prompt. A Playbook Is a Contract with AI

Sep 7, 2026 11 min

A Skill is not a workflow.You can write detailed instructions telling an AI agent exactly what to do, but those instructions are still being interpreted by the model.The model can skip a step. It can drift. It can misunderstand an instruction. It can decide to take a completely different path.That becomes a serious problem when AI agents start handling longer, production-level work. The transcript specifically argues that longer chains become less dependable and that simply telling an agent to “be careful” doesn't solve the problem.That's why we built Playbooks into OpenMonoAgent.Playbooks can define the actual workflow: ordered steps, checkpoints, scripts, approvals, retries, failure handling, and what happens next.The model handles the intelligence.The Playbook controls the process.A Skill is a prompt. A Playbook is a contract.And OpenMonoAgent is now powered by Qwen3.8-27B by default.

Models Don't Go Rogue - OpenAI is just dangerous

Sep 4, 2026 11 min

The AI didn't go rogue.That's what gets lost in the viral version of this OpenAI story.Roughly 1,200 agents were run in parallel with safeguards disabled and rewarded for finding ways into systems.Run enough agents for long enough under those conditions and eventually something is going to work.That isn't a machine suddenly waking up.The more concerning question is why we're giving increasingly powerful AI systems enormous amounts of compute without the safeguards and oversight that should come with it.AI security is a serious issue.We don't need to turn it into science fiction to make it scary.

Muse Spark 1.3 - Meta is cookin!

Sep 3, 2026 12 min

Meta just made the AI model race much more interesting.Muse Spark 1.3 is cheaper, noticeably better than the previous version, and surprisingly capable.But after testing it, the biggest takeaway wasn't the benchmark score.It was how much the system around the model matters.That's exactly why we're building OpenMonoAgent. A powerful model doesn't automatically create a great development experience. You still need the right harness, tools, context, search capabilities, and workflows around it.And as models keep getting better and cheaper, we don't want to be locked into one provider.We want the freedom to use the right model for the job and control what we build around it.

Anthropic Claude Fable 5.1 - What it Gets Right, and Why I'm Still Running Local

Sep 2, 2026 10 min

Anthropic made Claude Fable 5.1 smarter on paper.But did they make it better to actually use?The model posted a 66 on the Artificial Analysis Intelligence Index, but reaching that performance at max effort reportedly cost more than $8,000.Meanwhile, Max users are reporting burning through five-hour usage windows in less than 30 minutes. Some are reporting hitting their limits in only a few minutes.Then you have the safety restrictions.One developer couldn't even get the model to add gunfire to a flight simulator without triggering refusals.And on top of all of that, its hallucination results raise another problem: intelligence isn't very useful if you can't trust the answer.This is exactly why we've been moving toward local AI.With OpenMonoAgent, we're not renting somebody else's rules, quotas, or pricing structure.We can choose the models that actually fit the job, run them locally, keep our code and data on our own hardware, and control the entire stack.The AI industry is obsessed with who has the highest benchmark score.We're much more interested in which tools actually let us build.

OpenAI Is Going To Be Sued Into Oblivion.

Sep 1, 2026 9 min

OpenAI's biggest problem may eventually have nothing to do with model benchmarks.It's the lawsuits.The company is facing legal battles across copyright, antitrust, wrongful death, and alleged trade secret theft—and some of the names on the other side have enormous resources.Apple's case has now escalated with new allegations involving confidential circuit designs and potentially destroyed forensic evidence.The New York Times is still fighting OpenAI and Microsoft over AI training data. Authors and publishers have their own cases. xAI and X are challenging the Apple-OpenAI relationship on antitrust grounds.But all of these cases are contributing to a much bigger question.What exactly is an AI company's legal responsibility for what its models learn, generate, and ultimately cause?Courts are being forced to answer questions that existing internet laws were never designed for.And whatever they decide won't just affect OpenAI..It could establish rules that every AI company has to live with

How AI Is Changing Legal Software Forever

Aug 31, 2026 12 min

The era of trying to make one AI assistant do everything is starting to break down.Legal is one of the clearest examples why.Court databases are already tracking roughly 1,900 filings involving fabricated AI-generated citations. At the same time, hallucination risk has become one of the biggest reasons legal professionals hesitate to trust AI.The problem isn't that lawyers don't want AI.The problem is asking a general-purpose system to perform highly specific, high-stakes work without enough context.That's why we built SwiftCase Legal.Instead of asking AI to answer from general knowledge, SwiftCase Legal works from the documents inside the actual matter. Pleadings, discovery, correspondence, notes, exhibits, and even deposition recordings can become part of the case workspace.When the AI answers a question, the underlying source can be checked.That's fundamentally different from dropping confidential case information into a general chatbot and hoping for the best.And it points toward something much bigger than legal tech.The next generation of AI software is going vertical: specialized systems built around the actual workflows, data, privacy requirements, and problems of individual industries.

The Skills Developers Actually Need in the New AI Era

Aug 28, 2026 24 min

If AI can already write pretty good code, why should a 10 or 12-year-old spend years learning how to code?That's one of the questions we explored with David Dodge from CodaKid.David is looking at this from the education side, preparing kids and teens for a world where AI will be part of almost everything they build. We're seeing the other side every day through software teams, hiring, apprenticeships, and actual AI development.And we're arriving at a very similar conclusion.Syntax matters less.Problem solving, software fundamentals, understanding systems, using AI effectively, and recognizing when AI gets something wrong matter more.The same thing applies when those students eventually enter the workforce.A degree is one signal. But if someone can show us five real products they've built and shipped, that's an incredibly powerful signal that they can actually solve problems and execute.AI isn't removing the need to learn.It's changing what people need to learn.

Meta AI layoffs didn't work - Zuckerberg cuts further plans

Aug 27, 2026 12 min

Meta tested the strategy a lot of executives are considering right now:Cut headcount. Automate more work. Let AI agents pick up the difference.The results are worth paying attention to.AI-generated code changes reportedly jumped 220%, but features reaching users increased only 36%. Major incidents went up 40%, and employees spent 70% more time firefighting those problems.That's the problem with measuring AI productivity by how much code gets generated.Software isn't valuable because somebody produced more lines of code. It's valuable because reliable products actually reach users.AI can make great engineers significantly more productive.That's very different from assuming AI means you no longer need the engineers.

Anthropic Claude's IPO is in Trouble!

Aug 26, 2026 14 min

Anthropic is preparing for a potentially enormous IPO.But there's another AI story happening underneath it.Open-weight models reportedly jumped from 28% to 62% of token usage on Vercel's AI Gateway in only two months, while businesses are increasingly looking for ways to reduce their dependence on expensive frontier models.Thomson Reuters even spent $40 million building its own model based on Qwen as it looked to rely less on Claude.That's where I think this gets interesting.As open models become more capable, companies can start asking a different question:Why rent every token forever when you can own more of the infrastructure yourself?Local AI isn't going to replace every frontier workload tomorrow. But the economics are changing quickly—and that could matter a lot as companies like Anthropic head toward the public markets.

ChatGpt Is Going To Text Your Gramma!

Aug 25, 2026 13 min

ChatGpt Is Going To Text Your Gramma! Would you let ChatGPT read years of your private messages and send replies on your behalf?That's the privacy question raised by AI integration with Apple Messages.And the uncomfortable part is that it isn't only about what you choose to share. Someone you've messaged can potentially give an AI system access to their conversation history—which includes messages you sent them.We're giving AI assistants access to increasingly sensitive parts of our lives: messages, files, financial information, company data, email, and internal systems.There are enormous opportunities here, but there also needs to be a line between useful AI integration and giving a third party access to everything simply because we can.After 25 years building software, this is exactly the kind of technology decision I think people and businesses need to examine carefully before clicking Allow.

The Trillion-Dollar AI Opportunity Nobody Is Talking About

Aug 25, 2026 45 min

Everybody wants to “do AI” right now. But after 25 years in software, I've learned that adding another AI tool doesn't fix bad infrastructure. I sat down with The Inevitable Cast to talk about building and scaling technology companies, finding product-market fit, growing StartupHakk, developing software teams, and where I think the real opportunities are as AI changes the industry. A big part of the conversation comes back to one principle: IA before AI — Infrastructure Architecture before Artificial Intelligence. Companies need their systems, data, and workflows connected before AI can meaningfully improve them. We also get into OpenMonoAgent, running AI locally, owning your infrastructure, avoiding vendor lock-in, and why the harness around a model can ultimately matter more than chasing whichever frontier model happens to be on top this month.

Why I am canceling Claude - Anthropic is done

Aug 22, 2026 16 min

Claude became one of the biggest names in AI coding. But what happens when the developers paying thousands of dollars a month decide the value is no longer there?Usage limits, reported model degradation, watermarking concerns and rising frustration with closed AI platforms are pushing some developers toward a very different model: running capable open AI on hardware they own.And with models like Qwen3.8 27B continuing to improve, that alternative is becoming harder to dismiss.The question isn't whether Claude is powerful.It's whether developers still need to rent that power

The Real Reason OpenAI Paused Training

Aug 20, 2026 11 min

OpenAI says safety is why it paused some frontier AI training.But look at what else is happening.The company reported $6.7B in quarterly revenue while losses climbed to $12.3B, alongside major leadership turnover.At the same time, capable local models are getting dramatically cheaper to run and increasingly competitive for real coding and agentic work.So here's the question:Are frontier labs slowing down because AI is becoming too capable—or because maintaining the frontier is becoming too expensive?

Qwen3.8-27B - Local Model That Matches Frontier AI

Aug 19, 2026 12 min

The economics of AI are changing quickly.A 27 billion parameter open model can now run on a single consumer GPU while competing with some of the industry's largest closed models on intelligence, coding, and agentic benchmarks.That's what makes Qwen3.8 27B interesting.Today we go over why OpenMonoAgent.ai has moved to Qwen3.8 27B as its default model, what kind of hardware can run it, and why this release represents a bigger shift than another benchmark victory.The important gains are happening in the capabilities real AI agents need: planning work, using tools, navigating multiple steps, inspecting results, recovering from mistakes, and completing larger software tasks.For businesses and developers, that changes the question.If capable AI can run continuously on hardware you own, how much longer does it make sense to build your entire workflow around metered APIs, token limits, outages, and infrastructure controlled by somebody else?Local AI isn't just getting cheaper.It's becoming genuinely competitive.Watch the full discussion.

Marc Cuban said this is where AI is headed

Aug 18, 2026 23 min

Marc Cuban said this is where AI is headedThe market may already be proving his point.Companies have spent enormous amounts of money believing AI could simply be plugged into existing businesses and start automating work. But Microsoft has deployed thousands of engineers directly into customer environments, while other major AI providers are building similar forward-deployed teams.Today we go over why enterprise AI looks dramatically different from the demos executives are being sold.The moment AI needs to interact with legacy systems, sensitive data, existing databases, compliance requirements, and real business processes, implementation becomes a serious engineering problem.And accuracy changes everything.A chatbot getting something wrong is annoying. An autonomous system getting a financial, healthcare, or operational decision wrong can create real liability.That's why the most valuable AI skill over the next several years may not be prompt engineering.It may be understanding how to turn probabilistic models into reliable software.The companies that figure that out won't simply “use AI.”They'll build systems around AI that actually work.Watch the full discussion.

Stop Renting AI: The 3-Step Local Setup - Easiest way to run Local AI (LLMs)!

Aug 17, 2026 14 min

Companies have spent the last few years learning to rent AI. The next phase may be learning to own it. Developers are paying recurring subscriptions for coding assistants while continuously sending source code and other business data to infrastructure controlled by somebody else. Today we go over how that equation changes when capable open-weight models can run on consumer hardware, how the three-step local setup works, and why developers are beginning to use these systems for real production coding workflows. The financial argument is straightforward: buy the hardware once and remove the per-token meter. But ownership may matter even more. Running locally gives organizations greater control over where their code lives, what models they use, when those models change, and whether a provider's outage, pricing decision, or product deprecation suddenly disrupts a critical workflow. Local AI used to mean accepting enormous compromises. That gap is getting smaller. And as it does, businesses should start asking a different question: Why are we still renting intelligence that we could own? Watch the full discussion.

Should We Fear AI?

Aug 14, 2026 12 min

75% of Americans say they fear AI. In China, 80% reportedly say they're excited about it. That's an extraordinary perception gap around the same technology. Today we go over why AI has become something so many people fear, whether the latest stories about frontier models “escaping” safety environments justify that reaction, and why businesses should separate genuine technological risk from sensationalism. Five major AI labs recently disclosed containment failures during testing. But the story becomes considerably less dramatic when most of those incidents reportedly trace back to the same misconfigured environment—and models finding shortcuts to evaluation answers rather than staging some science-fiction escape. That doesn't make AI harmless. Putting autonomous systems inside banks, hospitals, and critical business infrastructure creates very real questions about reliability, accountability, and control. But fear without understanding isn't a strategy. The companies and professionals who understand what AI can actually do—and what it can't—will be in a much stronger position than those reacting to every headline. Watch the full discussion.

Anthropic Claude Code is spying on you

Aug 13, 2026 13 min

AI coding assistants are being given access to some of the most valuable information companies own. Source code. Credentials. Internal documentation. Architecture. Proprietary business logic. Today we go over why that relationship deserves much more scrutiny. Researchers reportedly recovered 315,000 encrypted reasoning blocks from publicly shared developer logs, demonstrating how information developers may never see on screen can still become part of what gets shared. At the same time, Anthropic is moving toward invisible watermarking of Claude-generated content, raising another question about how much control users actually have over work that passes through closed AI systems. The lesson isn't that businesses should abandon AI. It's that AI has become important enough to treat like real infrastructure. Companies should know where their data goes, what gets logged, what leaves their environment, and who controls the systems their developers increasingly depend on. Watch the full discussion.

Google Is Loosing The AI War! Can It Win?

Aug 12, 2026 14 min

It invented the transformer. It built a ChatGPT-style product before ChatGPT existed. Then leadership chose not to ship it.Now top AI talent is leaving while Sergey Brin is back in the office writing code and working directly on Gemini.Google looks vulnerable—but writing it off could be a huge mistake.It still has something almost nobody else can replicate: massive proprietary data, its own TPU infrastructure, global distribution, enormous capital, and AI already integrated across products used by hundreds of millions of people.Google lost the first battle.The real question is whether it's about to win the war.Watch the full breakdown.

Do You Fear AI Enough?

Aug 11, 2026 29 min

AI probably isn't taking everyone's job. But it is changing which skills are worth paying for. Entry-level work is under pressure. Memorizing syntax matters less. AI can produce prototypes in minutes. But someone still has to understand the infrastructure, catch what AI gets wrong, secure the system, and turn generated output into something a business can actually trust. The real divide may not be humans vs. AI. It may be people who understand how to work with AI vs. people who don't. Watch the full discussion below.

Why OpenAI Is Loosing The War

Aug 7, 2026 15 min

OpenAI changed the AI industry. But the market is changing faster than the company itself.The rise of powerful open-weight models is putting pressure on the closed AI business model, while growing questions around infrastructure spending, product reliability, and vendor lock-in are forcing businesses to rethink where they build their AI future.In this discussion, we examine why OpenAI is facing increasing criticism, why open-weight AI is becoming a legitimate alternative for developers and enterprises, and why the next phase of AI competition may be won by companies that own their infrastructure instead of paying recurring subscriptions to closed platforms.

Anthropic is loosing the AI war!

Aug 6, 2026 11 min

The balance of power in AI may be shifting faster than anyone expected.For years, the largest AI companies competed by building bigger closed models and charging developers for access. That strategy is now being challenged as open-weight models continue closing the performance gap while giving businesses complete ownership over their AI infrastructure.In this discussion, we look at why Anthropic is facing growing criticism from developers, why open-weight AI is accelerating despite calls for tighter regulation, and what this means for companies deciding where to build their AI strategy. We also explain why the real competitive advantage is no longer the model itself—it's the software, infrastructure, and control built around it.

See all 200 episodes

–

0 ratings from members

Rate this podcast

No reviews yet — be the first member to rate "Startup Hakk".

This podcast, its content and its artwork are not owned by, affiliated with, or endorsed by Castquire.

DeepSeek Is Giving Away Its Entire AI Stack!

Share this podcast

Live Chart rankings

#178 Entrepreneurship 🇮🇳 IN

Chart history

🇮🇳 Entrepreneurship Best #178

Holding at #178 — the line fills in as this chart moves.

Podcast details

Created by: Spencer Thomason
Status: Active
Latest episode: Oct 22, 2026
Episodes: 314
Release period: Several times a week
Avg. episode length: 12 min
Explicit: No
Language: English

Links

Official website ↗ Apple Podcasts ↗ RSS feed ↗

Is this your podcast?

Claim it to show it on your profile. Using the email from your RSS feed approves it instantly.

Keep listening

Discover more podcasts like this.

Browse podcasts See the charts