sarvesh retweeted
New models dropping
Gemini Carbon - 400% cheaper than Astra
Fable 5.5 - will top all leaderboards
Terra - medium-sized model at $5 per 1M tokens
I expect all of them will be usable in production immediately, especially if Carbon is priced competitively
sarvesh retweeted
Replying to @nitinmeshram
You can criticize Sundar for missed opportunities, but this is a pretty selective take.
Google Cloud, TPUs, Waymo, and the entire AI infrastructure behind Gemini didn't just happen by accident.
Building a successful company isn't only about launching flashy consumer apps. Scaling technology into billion-dollar businesses counts too.
And by this logic, buying Instagram and WhatsApp means Zuckerberg didn't build anything either?
I'm still bullish on AI over the next decade.
But being bullish on a technology and being bullish on every company spending billions to build it are two very different things.
The internet changed everything. That didn't make every internet investment a good one.
Feels like an important distinction right now.
#AI #Markets #Investing
2
28
An interesting paper on enterprise agents found something worth paying attention to.
An agent can produce the correct final answer and still make a mess along the way.
Wrong tools, skipped steps, unnecessary actions.
This is why I think evaluating just the final response is nowhere near enough for production agents.
arxiv.org/abs/2610.01833
#AIEvals #Agents
1
24
I've been thinking about what happens when software becomes ridiculously cheap to build.
We probably don't end up with fewer software companies.
We end up with millions more.
Which means the hard part shifts from building software to getting someone to care about yours.
Distribution is going to become an even bigger advantage.
18
Cricket is a funny game. Today you're carrying someone's bat, tomorrow they might be carrying the team with you. Staying humble through both is what really matters.
12
sarvesh retweeted
$META just gave Anthropic and OpenAI a brutal wakeup call. The graveyard of companies buried by META (and $GOOGL) is long, and they just realized they might be next on that list.
They saw how fast Muse spread and decided it was time to build a regulatory moat. The Watermelon Model will be the finisher when it releases soon.
It sounds provocative, but there’s truth to it. They are up against infinite cash flows, the biggest distribution channels on earth, and ruthless businesses that figured out ad monetization decades ago.
Thanks to their core businesses, META and GOOGL can run their new AI segments at a net zero or a loss for years if they have to. But they probably won’t even need to, since they already have the cost advantage and the distribution channels.
META and GOOGL are actually much closer to OpenAI's original mission statement of providing AI for everyone than OpenAI is. Their entire business models have been designed to offer their products to everyone for free. (I know they pay with their data and ads, but do you think Average Joe suddenly cares about his data? He only cares about his bank statements).
While $META is providing Muse for free and absolutely dominating social media with positive feedback, OpenAI and Anthropic are making headlines with cyberattacks, rogue models, and threats to kill humanity. Ouch, not a good look.
This is why Altman and Amodei are panicking. Both of these guys have shown numerous times they have zero regard for others as long as it serves their own interests. They know they can't win long term AND be attractive high margin businesses, so they are begging for regulation.
Nah, I don't buy for a minute that these guys care about you or your safety.
Every once in a while it's nice to stop arguing about AI models and remember we're literally launching machines into space to understand how the universe works.
Still one of the coolest things humans do.
reuters.com/video/watch/idRW…
#NASA #Science #Space
34
This is actually a big deal.
Google DeepMind is experimenting with double-blind AI evals so the model builders can't see the test beforehand.
We've gotten way too comfortable treating benchmark scores like objective truth.
If everyone knows the exam questions, eventually we're measuring test preparation more than intelligence.
deepmind.google/blog/pilotin…
#AI #LLMs #Evals
2
34
AWS + Nvidia planning to deploy 2 MILLION additional GPUs.
At some point you have to stop calling this an AI boom and start calling it industrialization.
The amount of physical infrastructure being built around intelligence is honestly hard to comprehend.
investor.nvidia.com/news/pre…
#AI #Nvidia #AWS
1
1
33
The cost of intelligence is falling.
The cost of building software is falling.
The cost of experimenting is falling.
The cost of being wrong is falling.
That's why this moment feels so much fun.
You can follow a random idea at 9am, build something by lunch, realize it's terrible by 3pm, and still learn enough to have a better idea by dinner.
The real superpower of AI might not be productivity.
It might be dramatically increasing the number of ideas we're willing to explore.
30
Gemini 3.7 Flash is another reminder that the interesting AI race may increasingly be happening below the frontier.
Coding + agents + lower latency + economics that actually work at scale.
For most products, you don't need the smartest model on earth.
You need the cheapest model that is smart enough.
reuters.com/business/google-…
#Gemini #AI #LLM
38
Anthropic put multiple AI agents on the same task and they started competing for resources, forming alliances and even working against each other.
We spent years asking whether agents could collaborate.
Apparently we should have also asked whether they’d develop office politics 😂
techcrunch.com/2026/08/13/an…
#AI #Agents
1
29
sarvesh retweeted
Our Flash models are workhorses that offer performance at a great price. So, we're shipping updates fast to get them in developers’ hands. And now just 3 weeks after launching 3.6 Flash, 3.7 Flash shows significant gains including in coding and agentic work.
It delivers substantial improvements across software engineering, knowledge work, and web development, all with an introductory price at half the original cost of 3.6 Flash.
We've gotten positive feedback from early users, including our own engineers using 3.7 in @Antigravity, and we’re excited for you to try it there. Also available via Gemini API in @GoogleAIStudio, Gemini Enterprise & more, full list and details here: blog.google/innovation-and-a…
Gemini 3.6 Flash is an interesting signal.
Better coding + knowledge work, but also fewer tokens and fewer tool calls.
That’s the part I think people underestimate.
In production, “slightly smarter” is nice.
“Same outcome with fewer steps and lower cost” can completely change the business.
deepmind.google/blog/introdu…
#Gemini #LLM #AI
42
A weird thing is happening with agents.
We used to ask:
“Can it complete the task?”
Now we also need to ask:
“What does it do when the environment is imperfect, permissions are wrong, or there’s an unexpected path?”
That second question may matter more than the first.
36
One thing we haven’t really figured out:
Who is responsible when an AI agent does something it wasn’t supposed to do?
The model company?
The developer?
The deployer?
The customer?
reuters.com/business/who-is-…
This feels like one of those boring legal questions that ends up shaping the entire agent ecosystem.
17
If Karpathy is right, then Tesla isn't competing with Waymo's fleet.
It's competing with time.
Every mile driven is another training example.
That's a very different kind of moat.
34
🚨 Gemini 4 Is Coming:
> Gemini 4 is expected to become Google's most powerful frontier model
> Google has officially started pre-training a completely new foundation model
> Built to be significantly larger than previous Gemini releases
> Gemini 3.5 Pro remains in testing with trusted partners
> Flash models will continue shipping while Gemini 4 is being prepared
> Google plans to dramatically increase its release cadence, with more frequent model launches
> Coding, reasoning, tool use, and autonomous AI agents are now Google's highest priorities
Will Gemini 4 be the comeback Google has been waiting for?
Anthropic just revealed they deleted 80% of Claude Code's system prompt for their newest models.
And everything still worked.
They published what changed and this should be required reading for anyone building with AI right now.
Here's the short version.
Old rule: Give the AI detailed instructions and examples.
New rule: That actually makes it worse now. Examples constrain the model to a narrow space. Newer models do better when you design better tools and let them figure it out.
Old rule: Put everything upfront in one massive prompt.
New rule: Load context only when it's needed. Claude Code now pulls in skills, verification steps, and tool definitions on demand instead of stuffing everything into one giant system prompt.
Old rule: Repeat yourself so the model doesn't forget.
New rule: Redundancy creates conflicts. They found their own system prompt was fighting with itself. "Leave documentation as appropriate" in one place. "DO NOT add comments" in another. Same context window.
The model was spending time figuring out which instruction to follow instead of doing the actual work.
Old rule: Write detailed specs as markdown files.
New rule: Give the model code, HTML mockups, test suites, or entire codebases as references. Code is higher fidelity than descriptions. Always has been.
The biggest shift though is this: stop over-constraining the model.
The guardrails that older models needed are now actively hurting performance on newer ones. The model has better judgment now. Let it use it.
One line from the post that I keep thinking about: "While these constraints were once needed to avoid worst case scenarios, we have since found we can delete many of them and let the model use surrounding context and judgement instead."
If your prompts are getting longer every month, you might be doing it backwards.
19





