0.000000	2.800000	 Tech It From Me is an independent and solo-produced podcast.
2.800000	7.440000	 Everyone's talking about OpenAI, Anthropic, and Google DeepMind,
7.440000	11.280000	 but while the headlines are dominated by billion-dollar labs,
11.280000	14.320000	 there's a quieter movement building in the background.
14.320000	17.440000	 Openweight, an open-source AI.
17.440000	19.840000	 It's fast-moving, it's innovative,
19.840000	23.920000	 and it might just be the biggest wildcard in the entire AI race.
23.920000	27.600000	 Welcome to the Tech It From Me podcast, I'm Mike Madole.
28.480000	30.880000	 Today, we're not talking about the closed-door labs
30.880000	34.320000	 with their secret sauce models and mystery training data.
34.320000	36.720000	 We're talking about the other side of AI.
36.720000	41.280000	 The stuff you can download, run, and even tinker with yourself.
41.280000	44.720000	 We're talking about projects like Mistral,
44.720000	47.760000	 Meta's LLaMA, and all the community-driven tools
47.760000	50.320000	 that anyone from a startup founder
50.320000	52.720000	 to a curious teenager can get their hands on.
53.520000	59.120000	 The question is, is open-source AI or Open-weight AI
59.120000	62.560000	 about to disrupt the big AI status quo,
62.560000	64.480000	 or will it always be playing catch-up?
64.480000	66.640000	 This is Tech It From Me.
66.640000	67.840000	 Let's go.
67.840000	70.320000	 This is the Tech It From Me podcast.
70.320000	74.480000	 Now, before we can talk about whether open-source
74.480000	78.640000	 or open-weight AI can compete with the big players,
78.640000	81.440000	 we need to be pretty clear what we actually mean
81.440000	87.360000	 by open-source because in AI, the term gets used pretty loosely.
87.360000	91.280000	 In traditional software, open-source is simple.
91.280000	96.320000	 You get the code, you can run it, change it, redistribute it,
96.320000	100.000000	 even sell it, all under an OSI-approved license.
100.000000	105.280000	 Linux, Apache, WordPress, that's true open-source.
105.280000	107.920000	 AI, though, is trickier.
108.640000	112.400000	 In most cases, when people say open-source AI,
112.400000	115.120000	 they really mean open-weight AI.
115.120000	120.160000	 That's when the creators release the actual trained parameters,
120.160000	123.280000	 the brain of the model, so you can run it yourself.
123.280000	127.920000	 But the license might still have rules about how you can use it.
127.920000	130.560000	 Take Mehta's Lama 3.
130.560000	134.960000	 Yeah, you can download it, run it locally, fine-tune it,
134.960000	136.800000	 and build apps on top of it.
136.800000	139.440000	 But it's not fully open in the purest sense.
139.440000	144.240000	 Mehta's license specifically blocks commercial use for companies
144.240000	147.680000	 with over 700 million monthly active users.
147.680000	151.120000	 In other words, sure, you can use this
151.120000	155.680000	 unless you're Google, Microsoft, or Apple.
155.680000	162.480000	 Then there's Mistral, a Paris-based AI company with fewer than 50 employees
162.480000	167.440000	 that's making waves, because they release models with minimal restrictions.
167.440000	173.280000	 They've said publicly they believe in open AI as a competitive equalizer,
173.280000	177.680000	 and their models are already being baked into all kinds of tools.
177.680000	182.960000	 And you can ignore the grassroots movement around projects like OLlama,
182.960000	188.960000	 which make it as easy as downloading an app to run a large language model on your own laptop.
189.600000	195.920000	 No cloud API keys, no reoccurring fees, just your own hardware, and the model.
195.920000	201.920000	 Now, compare that to the closed model world, OpenAI's GPT-40,
201.920000	207.120000	 Anthropic's Cloud 3, or Google's Gemini Ultra.
207.120000	210.880000	 You can use them, but only through their platform,
210.880000	213.840000	 and you'll never see their weights or training data.
213.840000	216.880000	 The companies say that's for safety reasons,
216.880000	222.000000	 but it also keeps them firmly in control of pricing, updates, and capabilities.
222.000000	226.800000	 So, when you hear open-source AI, it's worth asking,
226.800000	234.160000	 do they mean truly open, like Linux, or open-weight with a few legal speed bumps?
234.160000	237.760000	 Because in this race, the details really do matter.
237.760000	243.760000	 So, why is open-weight and open-source AI suddenly so hot right now?
244.720000	252.160000	 I'll give you four reasons. First, cost. Every API call to a closed model adds up.
252.160000	258.080000	 If you're a startup running GPT-40, your monthly bill could rival your rent.
258.080000	262.800000	 With open models, you can run them locally on a decent GPU,
262.800000	270.480000	 no pre-query fees, no unexpected invoices. Secondly, it's the speed of innovation.
271.840000	276.800000	 Close companies release updates every few months. In the open-source world,
276.800000	282.560000	 improvements can happen daily. Developers across the globe are fine-tuning models,
282.560000	287.360000	 adding features, fixing bugs, and sharing those improvements instantly.
287.360000	290.640000	 Thirdly, customization.
290.640000	296.640000	 Close models are like renting an apartment. You can decorate, but you can't knock down a wall.
296.640000	301.200000	 Open models are like owning the building. You can rip out the kitchen,
301.200000	307.200000	 add a new floor, turn it into a coffee shop. In industries like health care or a law,
307.200000	312.800000	 that flexibility is truly gold. And fourth, privacy and control.
312.800000	318.800000	 For some organizations, especially in finance, defense, or government,
318.800000	324.240000	 sending sensitive data to a third-party API is truly a non-starter.
324.240000	328.720000	 Running a model in-house keeps everything in your own environment.
329.520000	335.360000	 It's kind of like the Linux moment for AI. Linux never killed windows,
335.360000	339.440000	 but it became the backbone of the internet and countless embedded systems.
339.440000	345.760000	 Open AI models could follow that same trajectory, not replacing the giants entirely,
345.760000	350.800000	 but powering an enormous share of the ecosystem in the background.
350.800000	356.560000	 Now, of course, there are some pretty big hurdles. The first is obviously compute.
357.360000	363.840000	 Training a frontier model, something that will compete with GPT-4, now GPT-5,
363.840000	371.760000	 or cloud 3.5 costs tens or even hundreds of millions in GPUs, energy, and talent.
371.760000	378.720000	 Most open projects don't have that budget, so they focus on smaller, more efficient models.
378.720000	384.400000	 Then there's the quality gap. Open models are improving quickly,
384.960000	391.920000	 but if you're doing complex reasoning, nuanced conversation, or highly factual answers,
391.920000	398.400000	 the top closed models still have the edge. There's also security and misuse.
398.400000	405.520000	 An open model can be fine-tuned for malicious purposes, generating deepfakes,
405.520000	409.200000	 creating phishing emails, or even writing code for malware.
409.920000	417.040000	 With closed models, companies can filter or monitor usage with open ones. The responsibility falls
417.040000	425.520000	 entirely on the user. And finally, corporate capture. Even open projects often rely on funding
425.520000	432.000000	 from Big Tech or Venture Capital. If the money runs out, or the investors want to pivot,
432.000000	438.560000	 that openness could vanish overnight. The AI world has already seen open initiatives
438.560000	444.800000	 quietly close up shop once there's a profit motive. So it begs to question,
444.800000	450.640000	 can open-weight and open-source AI really stand toe-to-toe with the closed giants?
450.640000	458.160000	 Personally, I see three potential futures. Number one, domination in niches.
458.160000	466.000000	 Open models become the default for specialized industries where customization, privacy,
466.000000	472.560000	 and cost control matter more than cutting-edge performance. Think legal document review,
472.560000	477.200000	 customer service bots, or industry-specific research tools.
477.200000	486.640000	 Number two is a permanent second place. Frontier models, the absolute best, stay closed forever.
486.640000	492.560000	 The cost, the talent pool, and the safety concerns keep them behind company walls.
493.520000	498.320000	 Open models keep up in some areas but never quite match the bleeding edge.
498.320000	504.320000	 And then thirdly, the hybrid era. This is the one I think we're headed towards.
504.320000	508.640000	 Businesses will use open models for everyday automation,
508.640000	515.760000	 internal tools, and cost-sensitive workloads. And then call on closed models for the moments
515.760000	521.520000	 where they need maximum intelligence, reasoning, or multimodal capability.
522.480000	525.760000	 And make no mistake, the big players already know this.
525.760000	531.360000	 Meta isn't releasing Lama just to be nice. They're playing a strategic game,
531.360000	537.440000	 keeping themselves in the AI conversation, building developer loyalty, and putting competitive
537.440000	544.560000	 pressure on open AI and Google. Right now, open-weight AI is still the underdog.
544.560000	548.720000	 But history is full of underdogs that change the rules of the game.
549.600000	556.880000	 Linux didn't kill Windows. Android didn't kill iOS. But both reshape their industries in ways
556.880000	562.880000	 that force the big players to adapt. Open AI models could do exactly the same,
562.880000	570.080000	 even if they never hold the performance crown. The real question isn't if open AI will compete.
570.080000	574.880000	 It's how it will force the closed giants to change their game.
576.240000	579.600000	 Tech It From Me is an independent and solo-produced podcast.
579.600000	581.067688	 [BLANK_AUDIO]