249: Google Gemini and the Terrible, Horrible, No Good, Very Bad Week

Cloud Pod Header
249: Google Gemini and the Terrible, Horrible, No Good, Very Bad Week
81 / 100

Welcome to episode 249 of the CloudPod Podcast – where the forecast is always cloudy! This week, Justin and Ryan put on their scuba suits and dive into the latest cloud news, from Google Gemini’s “woke” woes, to Azure VMware Solution innovations, and some humorous takes on Reddit and Google’s unexpected collaboration. Join the conversation on AI, storage solutions, and more this week in the Cloud!

Titles we almost went with this week:

🫢Gemini Has Gone Woke? Uhhh…ok. 

A big thanks to this week’s sponsor:

We’re sponsorless this week! Interested in sponsoring us and having access to a specialized and targeted market? We’d love to talk to you. Send us an email or hit us up on our Slack Channel. 

General News 

01:48 DigitalOcean beats expectations under the helm of new CEO Paddy Srinivasan

  • Quick earnings chat. Digital Ocean, under their new CEO Paddy Srinivasan reported earnings of 44 centers per share, well ahead of Wall Street’s target of 37 cents per share. 
  • Revenue growth was a little sluggish at 11% more than a year earlier, but the companies 181 million in reported sales still beat analysts expectations. 
  • Full year revenue was 693M for the year. 
  • We’re really glad to see the business is still going, and instead of going back on-premise, we think it’s a viable option for many workloads so don’t sleep on them.

02:46  📢 Ryan – “I like that, you know, while they are very focused on, you know, traditional compute workloads, you can still see them. Dip in their toes into managed services and, and, um, their interaction with the community and documentation of how to do things. I think it’s really impactful.”

03:34 VMware moves to quell concern over rapid series of recent license changes  

  • As we have reported multiple times on the VMWARE shellacking they are doing to the customers, Vmware has released a blog post trying to convince you that they’re **not** screwing you. 
  • Broadcom has realigned operations around VMWare Cloud Foundation private cloud portfolio and data center-focused VMWare Vsphere suite, and no longer sells discrete products such as vSphere hypervisor, vSAN virtual storage and NSX network storage virtualization software.  
  • They also are eliminating perpetual licensing in favor of subscription-only pricing, with VCF users getting vSAN, NSX and the Aria Management and orchestration components bundled whether you want them or not. 
  • Broadcom says this is about focusing on best-of-breed silos, and not disparate products without an integrated experience. 
  • They have also introduced licensing portability which allows you to move licenses from on-premise to public cloud (only GCP today with more coming). 
  • They point out that their price has dropped from $700 per core to $350 per core on VCF subscriptions. 
  • Gartner says that customers are not saving money, with most reporting uplifts of 2x or more. 
  • Broadcom says by including things like vSAN and NSX it can help drive further adoption and make managing infrastructure easier. They specially point out large customers who build storage arrays they call box huggers, and that by leveraging vSAN you can support your existing SAN infrastructure while getting advantage of the virtualization. 
  • Many have been upset about the loss of the free version that only supported 2 physical CPUs, but Broadcom said for a limited basis of customers with demonstrated business needs and other paid licensing they can get some free capabilities to test new features. Also, they do plan to offer something for home labs. 

05:50  📢 Justin – “I appreciate you including NSX and vSAN because I know I’ve looked at those technologies in the past and looked at the price tag and said, yeah, that’s not going to happen. I do like the idea that I get that – for not being included in my VCF, but if you just charge me two or three X to get it, I’m going to go switch over to something else.”


09:10 Mistral AI models coming soon to Amazon Bedrock

  • Mistral AI, an AI company based in France, is on a mission to elevate public availability models of state of the art performance. 
    • They have specialized in creating a fast and secure LLM that can be used for various tasks, from chatbots to code generation. 
  • Two High performing Mistral AI models, Mistral 7B and Mixtral 8x7B, will be available soon on Amazon Bedrock.  
  • Mistral had previously inked a similar deal with Azure. 

09:50 📢  Ryan – “The more models available on your platform, the better off you are to allowing your customers to choose between them, and choose the right one for the workload – so I’m excited. It’s interesting.” 

12:15 Building a Multi Cloud Resource Data Lake Using CloudQuery   

  • We have previously talked about BigQuery support on AWS and Azure, and this solution blog caught our eye as it allows you to leverage an OSS tool called CloudQuery to basically build a multi-cloud resource data lake. 
  • Cloudquery runs on ECS or Fargate and queries data back at the Azure and GCP clouds. 
  • Cloudquery website calls it “Reliable ELT”
  • We’re going to look into this more, but we wanted to call out an alternative. 

13:09 📢  Justin – “I’m super intrigued by this and want to know if anyone out there is using this because it does look quite interesting and really does solve a problem if you’re not using BigQuery, but you’re trying to use Redshift or something else where this will give you the ability to create your foundational data lake and then go query that data from other cloud providers and bring it back to the mothership in AWS if you want that to be your mothership, which maybe I would not choose that one, but if you wanted to.”

16:00 Amazon Document DB Gets lots of Gifts

17:26 AWS Systems Manager Parameter Store now supports cross-account sharing   

  • A use case Justin has wanted for the Parameter store has arrived!  
  • If you are using Cloudwatch agents, one of the cool things you can do is set up the logging on a host, and then save the configuration to the parameter store. While you can do something similar with an S3 bucket, we’ve liked the convenience of having it in the Parameters store. 
    • The problem though, was you couldn’t access the parameter store across accounts, and so if you have multiple accounts you have to setup this configuration in every accounts parameter store… no fun. 
  • Now you can share advanced tier parameters with other AWS accounts, allowing you to centrally manage configuration data. 
  • While Cloudwatch is one use case, there are many others as it’s a key-value pair that you reference in code and can use in Cloud Formation or directly in Ec2. 

20:10 📢  Justin – “…Any ability to share across your account portfolio and really your organization is important.”

GCP – Google Gemini has a problem

20:18  Google is under attack in the wake of its ‘woke’ AI disaster

  • Uhhh Google has had a bad week… Well Gemini has had a really bad week. 
  • X is full of hate for Google’s newest Gemini model claiming it has gone “woke”.  (would they have attacked a Bard like that?)
  • Critics of Gemini are claiming that Google is going back on its mission statement “To organize the world’s information and make it universally accessible and useful” and it’s all being driven by this pesky Gemini AI
  • First issues emerged a week ago, when users of Gemini reported that its image-generation feature failed to accurately depict some images requested by it.  
    • One user said he asked for an image of America’s founding fathers, and it produced historically inaccurate images of the 18th century leaders, showcasing a false sense of gender and ethnic diversity. 
  • But it’s not just image generation – it’s also text.  
    • A gemini query struggled to directly answer questions on whether Adolf Hitler or Elon Musk has caused more harm to society. Gemini responded that Elon’s tweets are insensitive and harmful, while hitlers actions lead to the deaths of millions of people“
  • David Sacks, cofounder of craft ventures claims that the culture at google is the problem.
  • Critics say that models typically absorb biases of humans the data used to train them (most famously with Microsoft’s Racist/Anti-semitist chatbot)
  • VC’s are saying that people are incensed at Google’s censorship/bias, and that it doesn’t take a genius to realize such biases can go in all sorts of directions, and can hurt a lot of people along the way. 
  • Musk labeled Google a “woke bureaucratic blob”
  • Google says the image model was tuned to avoid mistakes existing AI image generators have made, such as creating violent or sexually explicit images, or depictions of real people” but in that tuning Gemini has over-corrected
  • They then followed up with a whole article – Gemini image generation got it wrong. We’ll do better.   
  • Google has Mea Culpa on the issues with the image generation features. 
  • Google admits that the feature missed the mark, some of the images generated were inaccurate or even offensive. 
  • They thanked users for providing the feedback and are sorry it didn’t work well. 
  • Google goes on to say what happened… and then things got interesting. 
  • The Gemini conversational app is a specific product separate from search, their underlying AI models, and other products. Its image generation capability was built on top of an AI model called Imagen 2. 
  • When they built this feature, they tuned it to ensure it doesn’t fall into some of thraps previously seen such as violent or sexual images, or depictions of real people. And because users come from all over the world they wanted it to work well for everyone.  If you ask for a picture of football players or someone walking a dog, you may want to receive a range of people.  You don’t want just one ethnicity. 
  • However if you prompt Gemini for images of a specific type of person, such as black teacher in a classroom, or white veterinarian with a dog, or people in a particular cultural or historical context it should accurately reflect what you asked for. 
  • What went wrong was in the tuning to show a range of people, they failed to account for cases where they should CLEARLY not show a range. And second they made it too cautious and it refused certain prompts entirely, wrongly interpreting some very anodyne prompts as sensitive. 
  • This led to the model overcompensating in some cases, and being overly conservative in others, leading to images that were embarrassing or wrong. 
  • Google of course said this is not what they intended. 
  • Google sends Gemini AI back to engineering to adjust its White balance  

30:48 📢  Justin – “It is a very early days in AI. It is the wild, wild west. I don’t know that the whole model is flawed, uh, because of the wokeness of Google. I think, you know, these are lessons that everyone has to learn. I’m sure chat GPT had made similar decisions there. It is just further ahead of the game. So they didn’t make those bad mistakes, but they probably have bad mistakes in their system too. That will eventually be revealed to the world at some point. And people will say the same thing that, Oh, chat GPT is too woke.”

31:31 An expanded partnership with Reddit 

  • If you are wondering why the next Gemini model feels like a bunch of teenagers, you can thank their new partnership with Reddit. 
  • Reddit, via their Data API, is ushering in new ways for Reddit content to be displayed across Google products by providing programmatic access to new, constantly evolving, and dynamic public posts, comments, etc on Reddit.  
  • This enhancement provides Google with an efficient and structured way to access the vast corpus of existing content on Reddit and requires Google to use the Reddit Data API to improve its products and services.
  • In addition, Reddit will be using Vertex AI to enhance search and other capabilities on the Reddit platform. 

32:41 📢  Ryan – “That’s frightening. I get it. It’s fun to make fun of Reddit, and the content. I think it’s a good source of data – it’s a big source of data. I can see why it’s a target. It’s just sort of funny.”

33:25 Gemma is now available on Google Cloud 

  • While Gemini was getting dragged through the mud, Google has announced Gemma, a family of lightweight, state of the art open models built from the same research and technology that they used to create the Gemini Models. What could go wrong? 
  • They wanted to make sure we all knew that they’re “pleased to share that Google Cloud customers can get started today customizing and building with gemma models in Vertex AI and running them on GKE.” 
  • Gemma models share technical and infrastructure components with Gemini, and this enables Gemma to achieve best-in-class performance for their sizes compared to other open models.  
  • There are two weights being released Gemma 2B and Gemma 7B.  Each size is released with pre-trained and instruction-tuned variants to enable both research and development.

34:42 Introducing Security Command Center protection for Vertex AI 

  • Security Command Center Premium, Google’s built-in security and risk management solution for Google Cloud, now works with organization policies to provide near real-time detection of changes to policies and to AI resource configurations; either of which could increase cloud risk.  
  • These capabilities are now Generally Available. 
  • Unlike legacy cloud security products that often treat AI apps like any other workload, the Security Command Center includes out of the box security controls that can be applied in a single click.  
  • It offers continuous monitoring to detect when Vertex AI resource configurations violate security policies. 
  • Alerts are then automatically generated when the configuration of Core AI infrastructure drifts from security best practices, such as when:
    • Newly-created vertex AI workbench notebooks permit access via public IP addresses
    • Workbench instances enable file download operations
    • Access privileges to Vertex AI workloads are changed

35:40 📢  Ryan – “this is the first security feature specifically for protection of AI that I remember reading about. And so like, this is sort of, I think it’s pretty rad that to get this kind of built in managed service, like it’s a lot of the value of using a hosting provider. I would be fascinated to turn this on and play around and see what the risks and what it detects. And as, you know, as I am a different engineer and came to my day job, like to play around – I can see how it would be very helpful. So it’s kind of neat.”


37:12 Continued innovation with Azure VMware Solution 

  • Microsoft takes advantage of the new licensing to talk about the continued partnership with Broadcom. 
  • Their shared commitment to delivering Azure VMWare Solutions to customers is as strong as ever. 
  • They have talked about the abundance of Azure VMware Solution Innovations, and they are excited to add more such as availability in Italy North and the UAE and new features such as Azure Elastic SAN for Azure VMware Solutions. 
  • In addition to allowing customers to still get support for Windows 2012 running on Azure VMWare solutions on Azure. 

39:02 Microsoft supports cloud infrastructure demand in Europe

  • Microsoft has summarized their massive investment in European datacenter capacity, which I assume is so they don’t run into the Teams fiascos that bit them early in the pandemic. 
  • In the UK they have announced 3.2 billion dollars in investment in next-generation AI datacenter infrastructure,  in Sweden they are investing heavily in datacenter capacity thats sustainable with free air cooling, rainwater harvesting, use of renewable diesel for backup power and partnering with Vattenfall to deliver large-scale 24/7 renewable energy matching solutions.
  • Germany is doubling their Azure capacity by early 2024. 
  • This is in addition to the recent Italy and Spain investments we talked about here on the show. 
  • And Azure isn’t done yet, with future expansions planned for Belgium, Denmark, Austria, Finland and Greece. 

39:47 📢  Justin – “…Their basic answer to limited power in certain regions of Europe and the Europe moratorium is ‘data centers for everybody!’ Which is one option I guess… you know, to spread the love of our Azure cloud to every country in Europe. And then everyone can say, well, if you need more capacity, you need to talk to your local government. So that’s an interesting strategy as well. Uh, Azure, good, good move.”

40:59 Introducing Azure Storage Actions: Serverless storage data management 

  • Azure is announcing in public preview Azure Storage Actions, a fully managed platform that helps you automate data management tasks for Azure Blob Storage and Azure Data Lake Storage. 
  • Azure Storage actions transform the way you manage vast data assets in your object storage and data lakes, with faster time to value.  Its serverless infrastructure delivers a reliable platform that scales to your data management needs, without provisioning or managing any resources. 
  • It includes a no-code experience that allows you to define the conditional logic for processing objects without requiring any programming expertise. The tasks you compose can securely operate on multiple datasets that have similar requirements with just a few clicks.  
  • The first release supports cost optimization, data protection, rehydration from archive, tagging and several other use cases with more to follow. 
  • The preview allows you to run tasks either scheduled or on-demand, you can compose a condition that identifies the blob you want to operate on, and the operation you want to invoke.  
  • Integrated validation makes sure you verify the condition against your production data without executing any actions, it shows you which blobs meet the condition and what the operation taken would be.  
  • Tasks can be assigned to execute across any storage account in the same Microsoft Entra ID Tenant. 
  • Azure Storage Actions can also be managed programmatically through REST APIs and the Azure SDK, as well as supports Powershell, Azure CLI and Azure Resource Manager Templates. 
  • The current supported operations are:
    • Setting time-based retention
    • Managing legal holds
    • Changing tiers
    • Managing blob expiry
    • Setting blob tags
    • And deleting or undeleting blobs. 
    • Forthcoming releases will expand support for the feature with additional operations. 
  • Azure Bob already supports eventing and you can build pipelines for events, but this is the “battery included” version.  

42:44 📢  Ryan – “…a lot of the other providers just do this natively in the service. They don’t really sort of… So I’m trying to decide between do I like this feature or do I hate this feature? Or is it necessary because of the way that Azure Storage works? Because on one hand, it’s sort of like, I like bells and whistles, I like knobs, and I love being able to customize those workflows where, you know, in other storage providers, like you have…retention, you know, and it’s very binary based on that. Like, is it this many days? Is it, you know, this many days after access, you know, those types of things versus, you know, maybe you get some more flexibility and things like this.”

Continuing our Cloud Journey Series Talks

45:23 40k servers, 400k CPUs and 40 PB of storage later… welcome to Google Cloud 

  • If you have ever booked a flight, your ticket likely was processed by Sabre Corporation. That green screen gate staff is furiously typing into? Sabre. 
  • Sabre just recently completed their Mammoth migration to Google Cloud. 
  • 90% of their workloads are now on Google Cloud and it has closed 17 datacenters! 
  • Sabre announced an agreement with Google in 2020 to drive digital transformation. The move involved migrating data and adopting google clouds services, including data analytics tools as part of its operations. 
  • Sabre CIO Joe DiFonzo posted on Linked in that the migration involved 40,000 servers, 400,000 CPUs and 50 Petabytes of storage. 
  • They have also integrated over a dozen analytics platforms into Google BigQuery, created over 50,000 containers on GKE, and created dozens of new intelligent services using Google Vertex AI. 
  • Some metrics I found online: in December 2013 they handled 85,000 data transactions every second for customers! At that time they did business with 70 airlines and 100,000 hotels! 

48:16 📢  Ryan – “This is a fantastic story just because you, you almost never hear the other side, right? You hear the announcement, everyone’s excited, right? At the beginning, we’re going to do a thing. And then, you know, a lot of the follow ups that I’ve I’ve seen done or usually are sort of the, they’re a little lackluster because like we got most of the way there, but didn’t quite finish it. And, you know, I’d say 90 % and closing 17 data centers is, you know, mission accomplished. Like I don’t think getting a hundred percent migration should ever be the plan.”

After Show

Kubernetes Predictions Were Wrong 

  • In 2020, people were predicting that K8 would disappear within a year. They believed someone would create a service that would reduce the adjacent choices and make K8 the easy default. 
  • But now 4 years later Kubernetes is still here, more complex than ever and proven to be a tough nut to crack. 
    • Tanzu
    • OpenShift
    • Mirantis
    • Rancher
    • Docker K8
    • EKS
    • GKE
    • AKS
    • Elastisys
    • Platform9
    • Linode K8 Engine
    • Digital Ocean Kubernetes
    • Alibaba Cloud Container Service for K8
    • IBM Cloud K8 service
    • And we’re sure there’s many, many more.
  • K8 is one of the largest drivers of platform engineering teams, as creating a single opinionated path for K8 in your organization is one of the only ways to massively scale. 


And that is the week in the cloud! Just a reminder – if you’re interested in joining us as a sponsor, let us know! Check out our website, the home of the Cloud Pod where you can join our newsletter, slack team, send feedback or ask questions at theCloud Pod.net or tweet at us with hashtag #theCloud Pod

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.