Mayuresh Bakshi

AI Engineering and Security. 2x founder. Early Engineering @AristaNetworks, @JuniperNetworks, @Cisco

[Xiaoyin Qu @quxiaoyin]

China’s AI playbook: kill OpenAI and anthropic with free great models. Make it free. Then use cheap electricity to export compute as well. Currently the blocker is chip but Hauwei would catch up soon. Imagine a world where instead of paying hundreds of billions to OpenAI and anthropic, you pay almost zero to similar level of intelligence with cheap cheap inference. What’s gonna happen?

[Nir Eyal @nireyal]

Confidence isn't knowing you'll succeed. It's knowing you'll figure it out.

Most people get this backwards. They think confidence is the feeling you get after you've already won — the certainty that the next thing will go your way because the last thing did.

That's not confidence. That's a winning streak.

Real confidence is quieter and more useful. It's the belief that whatever shows up, you'll find a way through it. Not because you have the answer in advance, but because you trust yourself to work the problem when it arrives.

This matters because the first version — "I know I'll succeed" — is fragile. The first failure breaks it. The second version — "I'll figure it out" — gets stronger every time something goes wrong, because every problem you work through is more evidence that you can.

Beliefs aren't facts. They're tools. And the belief that you'll figure it out is one of the most useful tools you can carry.

[Nikesh Arora @nikesharora]

Woke up in London to all the conversation of Chinese Cybersecurity models getting to Mythos like ability with agent swarms. Swarms that can explore vulnerabilites, determine attack paths and potential fixes in addition to persistent red teaming.

Happened just under 3 months, faster than my optimistic estimate. Expect that in a few weeks there will be more widespread capability. Highly likely that we will get US models released from bans faster with a promise of better hygiene.

What does it mean for the rest.
1. Test your own code!
2. Validate your vendors, ensure they are doing the same.
3. Start evaluating direct and virtual patching approaches to ensure open source is protected.

From a longer term perspective, we will need to ensure better security posture, no, no misconfigurations, robust platform products which can react swiftly, and a culture of constantly testing the enterprise with the most recent tools out there.

[Mayuresh Bakshi @bakshim]

In mature, deployed systems the Maintainer and Sweeper roles rightly take center stage to protect stability. The practical question becomes how a Prototyper can still run fast, low-risk experiments on top of substantial legacy code (via flags, abstractions, or sandboxes) without creating the very technical debt the later roles then have to pay down. Curious how teams are making that Prototyper-to-Builder transition reliable at scale.

[Mayuresh Bakshi @bakshim]

@petergyang Both can co-exist, doesn’t have to be one vs other, it depends on the consumption use case IMO.

[Mayuresh Bakshi @bakshim]

@levie @paulg When cyber AI hits open availability, security tooling economics change fast. Expect a wave of composable agents for continuous verification rather than periodic scans, worth exploring how that integrates with cloud-native infra at scale.

[Mayuresh Bakshi @bakshim]

When cyber AI hits open availability, security tooling economics change fast. Expect a wave of composable agents for continuous verification rather than periodic scans, worth exploring how that integrates with cloud-native infra at scale.

[Aaron Levie @levie]

It should be 100% obvious that there will soon be mythos level models on cyber security that are open and available to anyone. As a byproduct of this, alternative tech stacks will emerge that also drive more economic value and control away from the US’s tech stack.

This is what should be considered when thinking through the gate keeping you want to have in AI. If advanced models will become open and available regardless, then by not allowing the release of models you’re neither more secure nor better off strategically.

So much of the regulatory approach to AI has to assume China can’t catch up, when all current evidence suggests they can and are. And further, hard to imagine a higher priority than winning in AI for China; so you’re basically betting against their long term ingenuity, talent and motivation. Seems like a bad bet.

So your options are either to create gates around your best models, which means you’re asymmetrically disadvantaging yourself, or you work to ensure you’re always at the frontier and driving the future architectures of AI.

[Mayuresh Bakshi @bakshim]

Super clear breakdown, thanks @brian_armstrong . On the routing layer: do you use a small dedicated classifier model (or rules/embeddings) for preprocessing, or is it fully prompt-based with another LLM? Curious about the added latency and how you measure routing accuracy to avoid sending the wrong model.

[Mayuresh Bakshi @bakshim]

AI overhype :)
When builders are still struggling to go from prototype to production, tweets like this make the disconnect painfully obvious.

[Bhavish Aggarwal @bhash]

Got on the vibe coding bandwagon and built a bunch of AI agents this week for @OlaElectric. Wow!

So many layers get built between the actual doers and the founder as the company scales.

Agents will take away all middlemen in a company who are only “managing people” and not doing any problem solving!

And the people actually building will be even more valuable 🫡

[Mayuresh Bakshi @bakshim]

Love the opinionated stack for consistent controls @jedwards_27. Once these AI-generated apps are live and being used/composed with other tools, what kind of continuous security monitoring or anomaly detection do you have in place? Things like unexpected data access patterns or behavior changes. Also curious to know more about any sandboxing used to host

[Mayuresh Bakshi @bakshim]

Apple moment for OpenAI.
Massive ~840mm² compute die + 6 HBM modules. Optimized for inference - custom silicon era for inference is here.
NVIDIA still leads in versatility, but specialized chips like this will drive the next cost curve down

[OpenAI @OpenAI]

We’ve designed and built our first AI chip: Jalapeño.

Designed from the ground up by OpenAI and brought to production with @Broadcom, Jalapeño is purpose-built for the LLM workloads powering ChatGPT, Codex, the API, and future agentic products.

Chips are foundational to the AI economy. Building our own expands our full-stack platform from products to models to infrastructure, and will help us scale intelligence, serve more people, and expand access to AI.

[Mayuresh Bakshi @bakshim]

Great advice to build long term career!

[Nir Eyal @nireyal]

Work on the things people can't take away.

Titles get reorganized. Companies get acquired. Markets shift, layoffs happen, the role you trained five years for can disappear in a Tuesday morning meeting.

But your judgment is yours. Your skills are yours. The relationships you've built, the things you've learned how to do under pressure — nobody can pull those back.

Most career anxiety I hear is really anxiety about the wrong asset. People are guarding the title, the office, the org-chart position. Those are on loan. The lender can call them in.

What's actually yours is whatever you've built inside yourself: the craft you've sharpened, the trust you've earned, the way you think about problems.

So when you're deciding what to spend the next hour on — or the next year — that's the test worth running. Is this making me more of something nobody can take away? Or am I just defending borrowed ground?

Spend your effort on the first one.

[Arnav Gupta @championswimmer]

Every piece of software I use which used to be originally produced with a lot of care has gotten shitty. Just to make a list from top of my head...

1. Starting with this site. I used to give an example of how the Twitter mobile app was epitome of saving list scroll state across app lifecycle and even app death, all the way back in 2016 when teaching mobile development to my students. Today, most tweets > 2 day old if I open, the replies do not load, I don't get notifications for DMs, and random parts of it don't work at random times.

2. MacOS which was once more polished than Windows on the UI and as hackable as Linux from inside out - now randomly freezes, has kernel panics, needs disabling needless safety features all they way from safe mode to get basics working or toning down the horrible glass UIs.

3. Spotify used to be one of my favourite products, having great offline-first experiences, seamless sync across devices, handover of songs midway between phone, desktop, car, etc. Now the app can't even load offline downloaded playlists properly when internet is down, sync almost never works, UI glitches, watch app can't figure out how to play on headphone, or when to sync from phone to watch.

4. Whatsapp - one of the most performant apps, with solid delivery rates even with as slow as 2G/EDGE internet, now actually has dead-end UI flows (when sending photos, trying to edit it can lead to an unknown state), message deliveries often don't work even on solid internet, and media uploads frequently need retries.

5. Microsoft's entire office suite which used to be a workhorse product - something so reliable, that non-tech people would never touch Google Sheets with a 10-foot pole and threaten to resign if they didn't have a proper desktop app license of MS Office. Now they push you towards the cloud versions which work way worse than Google Workspace, and have add tons of React UI elements in the Desktop apps that makes then visibly slow and janky and large Excel sheets even crash sometimes.

Most of these were on the trajectory of enshittification before wide-scale agentic coding or Claude-driven development was even all that common.

The entire industry is in a phase where everyone is just building things because it is their job, and the era of care, and sincere craftsmanship of products has mostly come to an end.

[Mayuresh Bakshi @bakshim]

Great thread @nikesharora. Cheap tokens unlock pilots, but deployment needs verifiable outputs.
How do you see FDEs evolving to deliver deterministic guardrails + explainability at scale? Will this shift model labs toward more “consulting-like” roles (owning workflows + edge-case training) rather than pure model providers? Curious if verifiable AI becomes the real Phase 2 moat over raw capability.

[Deedy @deedydas]

Most software engineers are facing an identity crisis bordering on depression.

As CTOs aggressively evangelize tokenmaxxing, a class divide ensues.

The lazy. The lazy push code. They don't write it. They don't manually test it. They don't even read it. They're on autopilot. See Jira ticket, prompt for task, submit code. Many of them are barely on their computer the whole day. A comment on the PR asking why they did this? The lazy ask AI. A Slack message? The lazy ask AI. Need to prepare for standup? The lazy ask AI. As long as it sounds enough like them and isn't detected. Some of the lazy are even overemployed, and work multiple jobs. The lazy smart ones get away with this, and even rewarded. After all, software engineering for the lazy is just a dance to convince your colleagues you're smart and hard working.

The craftsmen. The craftsmen are tired. Very tired. 15 PRs in queue. Slack blowing up. The entire burden of review falls on the craftsman. The burden of understanding. They try. They work their way through the code, thoughtfully commenting to improve what ships. The response? A lazy: "That's a clever idea! You're absolutely right." with an incorrect change. It's fine, the craftsman says. I can fix them. They write a doc urging his colleagues to be better. The next day? 20,000 line PR to review. Day after day, their workload grows. Bugs seep into production. No one seems to care. Another round of AI is thrown at it. Their animosity to their colleagues rises. Eventually, they give up. It's just not what it used to be. The craft they loved is dead. They eventually wake up, a lazy.

This isn't all companies. Many companies are genuinely more productive, adopt the right set of principles and practices around AI development and have highly talented teams that trust each other. It tends to happen in bigger companies that are 10+yrs old with a higher talent variance. But it happens. A lot.

[Ben Carlson @awealthofcs]

It's gonna be weird now that a good labor market might be bearish for the stock market

Bad labor market = AI works

Good labor market = Is AI worth it?

[Aaron Levie @levie]

CEOs are uniquely prone to AI psychosis because they’re sufficiently distant from the last mile of work that still has to happen to generate most value with AI.

So when they play with AI, they see the happy path results, often not considering the next 10 or 20 things that have to happen to get sustainable results from agents.

“Look I made this awesome product prototype”. Yes but you didn’t have to review the code before it went into production and fix a bunch of issues.

“Look I generated a contract”. Yes but you didn’t verify all the terms before it goes out to the counterparty and didn’t have to wire up all the past contracts to work with.

The best thing you can do as a CEO is to use AI a *ton* to figure out the real implications of agents in the enterprise, and come out the other side with an appreciation for both the upside and the real work that goes into them.

[Mayuresh Bakshi @bakshim]

@chamath What are the chances that 8090 also takes the same route when you have enough data? Either way they are cooked :)

[Aaron Levie @levie]

One of the best things students and colleges can do is not bail on learning and teaching the fundamentals of any given domain. AI will trick you into thinking you don’t need to go deep in a particular area, but that’s wrong.

The expert with AI is always going to be far more capable than the novice. Those that can steer AI agents properly, figure out how to evaluate their work, fix their mistakes, and incorporate their work into a workflow will always be the most potent users of these tools.

The experienced software developer that’s built and scaled complex systems using agents outrun someone just vibe coding. The designer that uses AI will build far better products and campaigns than anyone else. The banker or analyst that understands financial models will be able to pull off far more with agents.

Despite some of the rhetoric in the valley that this is less implement now, that couldn’t be further from the case. Don’t give up on going deep in your craft.

[Mayuresh Bakshi @bakshim]

AI is enabler for attackers too. Securing the work of AI and making sure enough guardrails are placed around AI agents is going to be super important

[Aaron Levie @levie]

The ultimate rate limiter on productivity gains from agents will be on critical stuff like security, compliance, governance, the ability to review the work of the agent, ensure that it’s compatible with regulations, and so on.

We’ve been living in a little bit of la-la land around how much software enterprises are going to ultimately want to vibe code themselves. The last 48 hours represents a good example of why you won’t take on every risk of every piece of technology in your enterprise.

There’s no free lunch with AI productivity. Companies will have the build up the systems, processes, and controls for ensuring that agents can’t run around and do anything they want on any data at any time.