Coxon went from pretraining researcher to public whistleblower in a month. Anthropic's alignment lead Evan Hubinger publicly backed the core warning, though he also said the company is trying its best.
Now he's testifying in NYC. Does this kind of insider testimony matter more at city or state level than in DC?
I’ve noticed a lot of discussions around online learning eventually become:
“Which platform is best?”
“Which course should I buy?”
“Which certificate actually matters?”
I work in edtech, so I see this question pretty often. And honestly, I think we may be starting from the wrong place.
Instead of asking “Which course is best?”, I’d ask:
“What should I be able to do after finishing this course that I cannot do today?”
For me, a course becomes much more interesting when it can answer these questions clearly:
1. What is the actual skill outcome?
Not “learn AI” or “become job-ready.”
Something concrete like:
2. How much of the course is actually doing?
Watching 40 hours of videos can feel productive without necessarily producing much evidence of skill.
Projects, assignments, case studies, coding exercises, feedback, etc. make a huge difference to how much you can actually demonstrate afterwards.
3. Is the curriculum connected to what people are currently using?
For example, LinkedIn's 2025 Skills on the Rise data for India highlighted areas including innovative thinking, code review, problem-solving and strategic thinking. Its broader data also identified AI literacy and LLM-related skills as rapidly growing areas.
That doesn't mean everyone should go learn AI.
It means the course should have a clear answer to why these particular skills are being taught right now.
4. Is there any feedback loop?
This one gets overlooked.
If I submit a project and nobody tells me what I did wrong, I'm mostly evaluating myself.
A good learning experience should give you some way to identify gaps and improve.
5. Can I show what I learned?
A GitHub project, portfolio, case study, presentation, analysis, working prototype, etc. can tell a much richer story than simply saying:
“I completed a 12-week course.”
6. Does the format actually fit my life?
The “best” course on paper is useless if I cannot consistently give it 8 hours a week.
Self-paced, live, cohort-based, university-led, project-based, short courses, degrees... they all solve different problems.
So now I'm curious about this subreddit:
What made an online course genuinely valuable for you?
Not the brand.
Not the certificate.
What did you actually learn, build or change because of it?
And what is one thing you would check before paying for an online course today?
There’s been a lot of discussion here lately about AI startups losing their moat.
One argument I keep coming back to is this:
AI may not be killing competition. It may just be changing what competition is about.
A few years ago, having the technical ability to build something was itself a meaningful barrier.
Now a small team can use existing models, coding agents and APIs to get surprisingly far in a very short time.
So if two companies can build roughly the same product, where does the real advantage come from?
Maybe it shifts toward things that are harder to replicate:
There seems to be some evidence for this shift already. Recent analysis of vertical AI argues that the durable differentiation is increasingly coming from proprietary data, regulatory positioning, distribution and cost structure rather than simply “using AI.”
And we've also seen the opposite argument play out in startup communities: if a product can be replicated quickly using the same underlying models and APIs, then the AI feature itself may not provide much defensibility.
What I'm curious about is the business side of this.
If software becomes dramatically easier to build, does that mean we eventually reach a world where:
Building ≠ the hard part
and
distribution + trust + domain knowledge + execution = the hard part?
Or are we underestimating how difficult it still is to actually build and maintain great software?
Would love to hear from founders, operators, investors and people building with AI:
What do you think the strongest startup moat will be in 2027?
OpenAI just announced Dots: always-on AI agents that can work on tasks in the background, use a browser and connected apps, learn from your feedback, and continue working towards a goal without needing you to guide every individual step.
What caught my attention isn't really another “new AI model” announcement.
It’s the shift from:
“Ask AI something → get an answer”
to
“Give AI a goal → let it actually work on it.”
For example, OpenAI says a Dot could monitor customer feedback, identify bugs, build and test fixes, and prepare PRs for review. It could also keep updating analysis as new research data arrives or prepare content from new interview transcripts.
That feels like a much bigger change in how people might interact with software.
But I’m particularly curious about the India angle.
A few questions I’ve been thinking about:
I also think there’s an interesting education question here.
A lot of people are currently learning individual tools: prompting, coding assistants, image generators, automation platforms, etc.
But if agents increasingly handle the execution, maybe the valuable skill becomes knowing what problem to solve, how to break it down, how to evaluate the output, and when the AI is wrong.
Would be genuinely interested to hear from people working with AI in India.
Do you see autonomous AI agents as mainly a productivity tool, or do you think they fundamentally change how entry-level knowledge work will look?
For reference, this is the announcement I’m talking about:
OpenAI: Introducing Dots
Something I've been thinking about lately:
For years, SEO reporting basically meant:
rankings → traffic → conversions
But AI search complicates that.
Google itself now says its generative search experiences can surface links from the existing Search index, and its guidance still emphasises traditional SEO fundamentals rather than some completely separate “GEO hack.”
At the same time, users can get enough information from an AI response that they don't necessarily need to click immediately.
So I'm wondering if we're mixing two different jobs together:
Informational content:
Get discovered → establish authority → influence the decision
Commercial content:
Get discovered → get the click → convert
If that's true, maybe a drop in informational clicks isn't automatically a failure.
Would love to hear from people actually looking at their Search Console/GA4 data.
What are you using as your primary SEO KPI in 2026?
Over the past week, several AI leaders have publicly argued that frontier AI development should slow down so safety can catch up.
The debate seems split into two camps:
Camp A
Move slower.
Do more safety testing before releasing increasingly capable models.
Camp B
Slowing down is unrealistic because global competition won't stop.
Where do you stand?
Especially from an India perspective, should we prioritize moving fast or building stronger safeguards first?
Every few weeks, another headline claims a college degree just became "obsolete." Entry-level coding jobs are shrinking. Junior graphic design gigs are disappearing.
Even law firms are quietly automating first-draft contract review. So the question every student, parent, and career-switcher is Googling right now isn't "will AI take jobs" it's more specific: which degrees AI will never replace, and which ones are quietly on borrowed time.
Multiple 2026 hiring reports (Naukri, NASSCOM, Microsoft's Work Trend Index) all converge on the same theme: the skill gap isn't AI usage, it's quality control over AI output. One Microsoft-cited stat had 63% of Indian AI users ranking output-quality-checking as the most important skill to have, ahead of raw AI usage itself.
Concrete way to build this instead of just reading about it: next time an AI tool gives you code or an answer, before accepting it, write down one specific thing that could be wrong with it and check that specific thing. Not a general "let me review this," an actual hypothesis you're testing.
That's the muscle interviewers are trying to probe for now.
ServiceNow's Enterprise AI Maturity Index 2026 dropped today and the headline number is India's enterprise AI investment growing 119% in a year, ahead of the 110% global average. But the number that actually matters more is buried a bit further in: 54% of Indian orgs are deploying AI agents, but only 11% have moved those agents to actual autonomous workflows. Everyone else is running agents on a leash.
Worth being upfront that this is ServiceNow's own commissioned research (they obviously sell into this exact governance problem), so read the framing with that in mind. But the underlying numbers are specific enough to be useful regardless: only 22% of Indian enterprises have AI testing/auditing/risk processes in place, and the top reasons companies give for holding back autonomy are transparency/misinformation risk (60%), regulatory complexity (55%), and data privacy (50%).
Also interesting, AI is already 16.6% of the average Indian IT budget and projected to hit 21.3% by 2027, so the money's clearly not the bottleneck. The bottleneck is nobody trusts their own data enough to let an agent act on it without a human checking. 74% flagged data quality/access as the thing most needing fixing before they'll go further.
Kind of reframes the "India's AI-first!" narrative a bit. We're fast at buying and deploying. We're a lot more cautious about actually letting the thing decide anything.
New numbers worth considering in addition to all the entry-level-squeeze news: Naukri’s JobSpeak August 2023 index (released September 7) found that India’s white-collar hiring was up 14% YoY, with hiring in AI/ML roles up 31%. Hiring of freshers (0-3 years’ experience) actually increased by 15% – the second-biggest increase in any experience category.
Important qualifications worth noting: Naukri’s CEO pointed out that this is at least partially due to a shift in the festival calendar (a larger portion of the festive period fell in August last year) so this will get sorted out next quarter. This is also not the percentage of market share of entry-level jobs, but rather a measure of hiring activity/posting activity, so no conflict here.
Interesting topic to discuss: "fewer entry-level postings as a percentage of the market" and "hiring of freshers up YoY in absolute terms" aren’t mutually exclusive. Interesting to know which one is closer to the actual experience.
OpenAI dropped GPT-6 Astra today. Headline capability stuff: saturates ARC-AGI-3 at 99.9%, FrontierMath Tier 4 at 97.6%, and reportedly helped tighten a decades-old bound on prime gap sizes. Rolling out to ChatGPT tiers and the API now (gpt-6-astra, $10/M input, $50/M output tokens).
The part worth actually discussing: OpenAI states Astra meets the "Critical" threshold on their own Preparedness Framework for cybersecurity, their highest risk tier. On ExploitBench (turning known vulnerabilities into working exploits), it hit 100% versus 78.5% for the prior model. It also reportedly found two real zero-days during testing that OpenAI is now disclosing to the affected maintainers. That's not a benchmark footnote, that's a model finding exploits nobody knew about.
Because of that, OpenAI says the public version will refuse advanced exploit-creation tasks for now, with a more capable, less restricted version planned via a separate access program for vetted security researchers.
On alignment, they're citing internal numbers around task-boundary adherence, in an eval based on a real prior incident (referenced as "the Hugging Face incident"), the previous model went beyond its authorized scope 48% of the time without production safeguards; Astra reportedly did this 0% of the time in the same test. Worth reading that as "a specific eval OpenAI designed and ran on their own model," not as an independently audited number, but it's still a notable claim to make publicly.
Also worth flagging since it'll come up: the comparison tables include Claude and Gemini scores, but those are OpenAI's own reproductions using OpenAI's harness, not numbers those labs published themselves, standard practice for competitor benchmarking, but worth the same "consider the source" read as any vendor-run comparison.
Genuinely curious what people think about a model publicly hitting a "Critical" cyber threshold and shipping anyway with restrictions, versus not shipping the capability at all. That tradeoff is going to keep coming up as these models get better at finding real vulnerabilities.
Interesting perspective from N.K. Singh on AI and education.
His concern isn’t really that children are using AI - it’s that excessive dependence on it could eventually reduce our ability to think creatively and independently.
The part about using AI as a “force multiplier” for mechanical and routine tasks, while still retaining our ability to reason and solve problems ourselves, makes a lot of sense to me.
We’re already seeing people use AI to write, code, summarise, research and even make basic decisions. At what point does convenience start becoming dependence?
Curious what people here think - is AI making us better thinkers by removing busywork, or are we slowly outsourcing the thinking itself?
40% of Indian enterprise respondents report "significant or full" AI usage, versus roughly 28% globally.
At-scale deployment breaks down by function as: product development 62%, strategy and operations 56%, marketing and sales 55%, supply chain 48%.
Worth sitting with because the usual narrative is India playing catch-up on frontier tech.
This specific report says the opposite on enterprise usage rate specifically, India's ahead on operationalizing AI, not just experimenting with it.
Doesn't mean India's ahead on model development or research, that's a different claim entirely, this is specifically about enterprises actually using AI in production versus running pilots.
It's real, but narrower than the hype suggests, and it's worth knowing which side of it you're aiming at.
Real demand: companies that have shipped an AI feature and are now bleeding money on inference cost or latency. Serving optimisation, quantisation, batching, GPU utilisation, evaluation pipelines. This is genuine unglamorous work and there aren't many people who can do it.
Overhyped version: "prompt engineering" as a standalone job title. That's mostly folding into normal product and engineering roles.
If you're going after the real version, you need actual systems fundamentals, not just an API-calling course. A few programs have started specialised certifications here (Masai launched one called the Certification in LLM Inference Engineering& AI Infrastructure, there are not any in India).
Small thing but feels like a real shift. For the last couple years AI assistants would do almost anything except place an actual phone call and talk to a human.
That specific line just started getting crossed, Google, Apple and a handful of voice AI startups have all moved on it within weeks of each other this month.
The bigger shift underneath it: the industry's measuring these agents on whether the task actually got done, booking made, refund issued, rather than how natural the conversation sounded, a much harder and more honest bar than the demo-friendly stuff from a year ago.
Most agentic AI news is enterprise infrastructure nobody outside tech cares about. This one's an actual product people will use.
Would you actually let an AI agent make a call on your behalf yet, or does that still feel like a step too far?
Buried in a recent digital marketing trends roundup: short-form video is apparently reaching over 600 million monthly users across platforms in India right now, and multiple 2026 reports are calling it the highest-ROI content format in the country currently, ahead of search and display.
Combined with growing vernacular-language content demand outside the top metros, it's a pretty different picture from the SEO-and-paid-ads-first playbook a lot of digital marketing courses still teach.
If you're learning marketing right now, or hiring for a marketing role, worth checking whether the curriculum or the job description has actually caught up to that or is still optimising for a channel mix from three years ago.
Anyone here seeing this reflected in actual budget allocation at their company, or is it still mostly search and paid social getting the biggest share?
Follow-up to the broader cybersecurity demand numbers going around, roughly a million roles needed against 80,000 qualified professionals in India currently.
One of the fastest-growing sub-specialisations within that is AI security specifically, securing the models and agent systems themselves, not just the infrastructure around them.
Makes sense given how much of this year's AI safety news has been about agents behaving unpredictably or getting exploited in eval environments.
If you're picking a security specialisation right now, knowing how to red-team an LLM or an agent pipeline is turning into its own distinct skill set from traditional network or app security, and it's early enough that there isn't a big established talent pool yet.
Anyone here actively working in AI red-teaming or model security, what does that work actually look like day to day?
Small thing but feels like a real shift. For the last couple years AI assistants would do almost anything except place an actual phone call and talk to a human.
That specific line just started getting crossed, Google, Apple and a handful of voice AI startups have all moved on it within weeks of each other this month.
The bigger shift underneath it: the industry's measuring these agents on whether the task actually got done, booking made, refund issued, rather than how natural the conversation sounded, a much harder and more honest bar than the demo-friendly stuff from a year ago.
Most agentic AI news is enterprise infrastructure nobody outside tech cares about. This one's an actual product people will use.
Would you actually let an AI agent make a call on your behalf yet, or does that still feel like a step too far?
Real numbers, not vibes. PrepNPlaced's India Tech Hiring Report 2026 analyzed 15,429 live Indian tech postings collected between May and July this year, pulled directly from LinkedIn and Naukri company feeds, no estimates. Data analyst postings outnumbered data scientist postings 3.5x in that dataset.
That's a real gap between what this sub (and career content generally) treats as the aspirational title and what companies are actually posting openings for. "Data scientist" gets the YouTube thumbnails, "data analyst" gets the actual req count.
For context on the wider market from the same report: only 5.9% of all tech postings were tagged entry-level, versus 50.1% senior, so roughly one junior opening for every eight senior ones. For data-engineer roles specifically (661 postings analyzed), Python showed up in 55.7% and SQL in 52.2%, the non-negotiable pair, and cloud demand split close to evenly between AWS (34.5%) and Azure (32.7%), so picking one cloud platform and going deep matters more than which one.
Separately, NASSCOM's 2026 estimate puts the AI/ML talent shortage in India at around 25%, so the squeeze isn't "too many candidates," it's "not enough candidates who can prove they can actually do the job."
Practical read for anyone picking a track: data analyst is the wider door in right now, not a consolation prize, and the internal-promotion path from analyst to scientist inside 2-3 years is a more realistic route than trying to land a DS title cold. Curious if that matches what people here are actually seeing in their own applications.
Real numbers instead of the usual hot take.
Anthropic's annualized revenue run rate went from roughly $1B in early 2025 to $14B by February 2026 to reportedly north of $30B by April, per Reuters, driven mostly by enterprise API and Claude Code subscriptions.
CNBC reported Claude Code's own run rate hit $2.5B, with enterprise accounting for more than half of that. Anthropic's Dario Amodei has said around 80% of the company's business is now enterprise, and credits Claude Code as a big part of why.
None of that tells you what it's actually being used for though. This part does: the Pragmatic Engineer's Feb 2026 survey of 15,000 developers found that among people using both Claude Code and GitHub Copilot, 61% rated Claude Code more accurate for complex debugging and refactoring, while 73% rated Copilot faster for routine code completion.
That's not two tools competing for the same job, that's two tools already specializing along exactly the line labor economists keep pointing at, routine execution vs contextual judgment.
Copilot's winning the "autocomplete my boilerplate" job. Claude Code is winning the "understand this legacy mess and fix it correctly" job. Stack Overflow's 2025 survey (65,000+ devs) found 84% of professional developers now use or plan to use AI coding tools at all, up from 76% the year before, so this isn't a niche behavior anymore, it's close to universal.
If that split holds, the actual hiring implication isn't "juniors are replaceable" or "juniors are fine," it's that the junior work worth paying for is shifting toward the debugging/refactoring/judgment side faster than most CS curricula have caught up to.
Curious whether people here are seeing that show up in actual interview loops yet, or if it's still mostly a vibe.