look like an engineer downloading Google=E2=80=99s Antigravity=
IDE on their work laptop without waiting for security review, or a teammat=
e using Claude Code from their terminal while the official tooling request =
sits in procurement limbo.
Shadow AI poses security threats to organizations=E2=80=94and expands the o=
verall attack surface for data breaches and sensitive information leaks. Ac=
cording to Netskope=E2=80=99s 2026 [Cloud and Threat Report](https://www.ne=
tskope.com/resources/cloud-and-threat-reports/cloud-and-threat-report-2026?=
utm_medium=3Demail&utm_source=3Dnewsletter&utm_campaign=3DAugust-2025), the=
three categories of data most involved in AI data policy violations over t=
he past year were source code (42%), regulated data (32%), and intellectual=
property (16%). For example, [Samsung discovered](https://mashable.com/art=
icle/samsung-chatgpt-leak-details?utm_medium=3Demail&utm_source=3Dnewslette=
r&utm_campaign=3DAugust-2025) that its proprietary semiconductor data was l=
eaked through internal engineering use of ChatGPT to check source code, opt=
imize code, and record meeting notes for a presentation.
Without the right guardrails and oversight, shadow AI will continue to outg=
row organizational tooling and systems, further putting data at risk. Wheth=
er you manage AI tools within an organization or use them yourself, address=
ing the risk of shadow AI keeps [business and customer data safe](https://w=
ww.digitalocean.com/resources/articles/ai-and-privacy?utm_medium=3Demail&ut=
m_source=3Dnewsletter&utm_campaign=3DAugust-2025&utm_source=3Dmarketo&utm_m=
edium=3Demail&utm_campaign=3Dai-newsletter-jan-2nd-edition&utm_content=3Dsh=
adow-ai-article).
How to keep shadow AI at bay
For developers, be an advocate, not a rogue agent. Test new tools on person=
al devices, document the productivity gains, and bring a case to your team =
lead or IT. You're more likely to get approval when you show up with findin=
gs instead of a request.
For leaders and IT administrators, rigid six-month procurement cycles are p=
art of the problem. Consider faster evaluation pipelines or sandboxed trial=
s, processes that let teams experiment safely without pushing them toward w=
orkarounds, so your team can stay competitive without compromising security=
.
AI trend reports and industry commentary
- Ads have landed in ChatGPT: [OpenAI](https://openai.com/index/our-approac=
h-to-advertising-and-expanding-access/?utm_medium=3Demail&utm_source=3Dnews=
letter&utm_campaign=3DAugust-2025) announced that it would start testing ad=
s on the Free and Go tiers, noting that a full-scale rollout is not yet pla=
nned.
- A joint partnership between Google and Apple: [Google and Apple](https://=
blog.google/company-news/inside-google/company-announcements/joint-statemen=
t-google-apple/?utm_medium=3Demail&utm_source=3Dnewsletter&utm_campaign=3DA=
ugust-2025) have announced a multi-year collaboration to build Apple Founda=
tion Models on Google's technology, reducing OpenAI=E2=80=99s competitive h=
old in the market.
Models on our radar
- [GLM-Image](https://huggingface.co/zai-org/GLM-Image?utm_medium=3Demail&u=
tm_source=3Dnewsletter&utm_campaign=3DAugust-2025): GLM-Image is an image g=
eneration model that adopts a hybrid autoregressive + diffusion decoder arc=
hitecture. In general image generation quality, GLM=E2=80=91Image aligns wi=
th mainstream latent diffusion approaches, but it shows significant advanta=
ges in text-rendering and knowledge=E2=80=91intensive generation scenarios.
- [Pocket-TTS](https://huggingface.co/kyutai/pocket-tts?utm_medium=3Demail&=
utm_source=3Dnewsletter&utm_campaign=3DAugust-2025): A lightweight text-to-=
speech (TTS) application designed to run efficiently on CPUs. Forget about =
the hassle of using GPUs and web APIs serving TTS models. With Kyutai's Poc=
ket TTS, generating audio is just a pip install and a function call away.
- [Flux.2 Klein](https://huggingface.co/black-forest-labs/FLUX.2-klein-9B?u=
tm_medium=3Demail&utm_source=3Dnewsletter&utm_campaign=3DAugust-2025): The =
FLUX.2 [klein] model family are our fastest image models to date. FLUX.2 [k=
lein] unifies generation and editing in a single compact architecture, deli=
vering state-of-the-art quality with end-to-end inference in as low as unde=
r a second. Built for applications that require real-time image generation =
without sacrificing quality.
- [GLM-4.7](https://huggingface.co/zai-org/GLM-4.7-Flash?utm_medium=3Demail=
&utm_source=3Dnewsletter&utm_campaign=3DAugust-2025): GLM-4.7, labeled "you=
r new coding partner," delivers major improvements over GLM-4.6 across core=
coding, reasoning, UI generation, and tool use, with strong gains in multi=
lingual and agentic coding and terminal-based tasks=E2=80=94scoring 73.8% (=
+5.8%) on SWE-bench, 66.7% (+12.9%) on SWE-bench Multilingual, and 41% (+16=
.5%) on Terminal Bench 2.0=E2=80=94while also supporting "thinking before a=
cting" for better performance in popular agent frameworks like Claude Code,=
Kilo Code, Cline, and Roo Code.
- [Agent-CPM Explore](https://huggingface.co/openbmb/AgentCPM-Explore?utm_m=
edium=3Demail&utm_source=3Dnewsletter&utm_campaign=3DAugust-2025): AgentCPM=
-Explore is an agent foundation model with only 4B parameters in its entire=
training and inference infrastructure. AgentCPM-Explore has successfully e=
ntered 8 classic long-horizon agent benchmarks, including GAIA, HLE, and Br=
owserComp. AgentCPM-Explore achieves SOTA performance at the same parameter=
scale and demonstrates its accurate deep research capabilities, effectivel=
y breaking the performance bottleneck for on-device agents.
Community tutorials and resources
- Agent skills are folders of instructions, scripts, and resources that an =
LLM can load to perform tasks. First introduced by Claude in 2025, this art=
icle [creates an Agent Skill](https://www.digitalocean.com/community/tutori=
als/how-to-implement-agent-skills?utm_medium=3Demail&utm_source=3Dnewslette=
r&utm_campaign=3DAugust-2025&utm_source=3Dmarketo&utm_medium=3Demail&utm_ca=
mpaign=3Dai-newsletter-jan-2nd-edition&utm_content=3Dagent-skills-tutorial)=
for a PDF parser.
- [LangSmith is a platform](https://www.digitalocean.com/community/tutorial=
s/langsmith-debudding-evaluating-llm-agents?utm_medium=3Demail&utm_source=
=3Dnewsletter&utm_campaign=3DAugust-2025&utm_source=3Dmarketo&utm_medium=3D=
email&utm_campaign=3Dai-newsletter-jan-2nd-edition&utm_content=3Dlangsmith-=
tutorial) for instrumenting, debugging, evaluating, and monitoring LLM apps=
to overcome these challenges and provide end-to-end visibility into your a=
gent's behavior. In this tutorial, we will go through a typical agent relia=
bility loop: instrumenting an agent with tracing, debugging agent runs auto=
matically and manually, evaluating its performance, iterating on prompts, a=
nd finally, monitoring the agent in production.
- A true step forward for open-source video generation, LTX-2 is undoubtedl=
y the most powerful video generation technology ever open-sourced. In this =
tutorial, we will walk through [how LTX-2 works](https://www.digitalocean.c=
om/community/tutorials/ltx-2-video-generation-audio-video-model?utm_medium=
=3Demail&utm_source=3Dnewsletter&utm_campaign=3DAugust-2025&utm_source=3Dma=
rketo&utm_medium=3Demail&utm_campaign=3Dai-newsletter-jan-2nd-edition&utm_c=
ontent=3Dltx-2-tutorial) by exploring its architectural pipeline, then show=
how to run LTX-2 on DigitalOcean Gradient using the ComfyUI.
- Vercel's slick Git-based deploys and edge functions have made it a fronte=
nd favorite, but it's not the only option. Compare [Vercel alternatives](ht=
tps://www.digitalocean.com/resources/articles/vercel-alternatives?utm_mediu=
m=3Demail&utm_source=3Dnewsletter&utm_campaign=3DAugust-2025&utm_source=3Dm=
arketo&utm_medium=3Demail&utm_campaign=3Dai-newsletter-jan-2nd-edition&utm_=
content=3Dvercel-alternatives-article), including DigitalOcean App Platform=
, with breakdowns of CI/CD workflows, scaling, and pricing.
- Need to store petabytes of unstructured data without blowing your budget =
on egress fees? Check out these [object storage solutions](https://www.digi=
talocean.com/resources/articles/object-storage-solutions-cloud-data?utm_med=
ium=3Demail&utm_source=3Dnewsletter&utm_campaign=3DAugust-2025&utm_source=
=3Dmarketo&utm_medium=3Demail&utm_campaign=3Dai-newsletter-jan-2nd-edition&=
utm_content=3Dobject-storage-article)=E2=80=94we compare S3 compatibility, =
pricing models, zero-egress options, and AI/ML pipeline integration across =
providers from hyperscalers to lean alternatives.
- Choosing a VPS provider shapes everything from deployment speed to long-t=
erm costs. [DigitalOcean vs AWS Lightsail](https://www.digitalocean.com/res=
ources/articles/digitalocean-vs-awslightsail?utm_medium=3Demail&utm_source=
=3Dnewsletter&utm_campaign=3DAugust-2025&utm_source=3Dmarketo&utm_medium=3D=
email&utm_campaign=3Dai-newsletter-jan-2nd-edition&utm_content=3Ddo-vs-ligh=
tsail-article) breaks down pricing transparency, egress fees, performance c=
onsistency, and support tiers to help you decide which platform fits your w=
orkload.
Video tutorials
[How AI models run on GPUs (CUDA vs. ROCm)](https://www.youtube.com/watch?u=
tm_medium=3Demail&utm_source=3Dnewsletter&utm_campaign=3DAugust-2025&v=3DqL=
MjoLIzkR4&utm_source=3Dmarketo&utm_medium=3Demail&utm_campaign=3Dai-newslet=
ter-jan-2nd-edition&utm_content=3Dcuda-rocm-video)
Ever wondered how your Python code actually runs on a GPU? This video break=
s down the AI compute stack=E2=80=94the layers of technology that transform=
your high-level PyTorch commands into calculations running on thousands of=
GPU cores. Whether you've heard of NVIDIA's CUDA or AMD's ROCm but never u=
nderstood what they actually do, this explainer will give you the complete =
picture from silicon to Python.
Product updates
[GPU Droplets accelerated by NVIDIA HGXTM B300 are coming soon](https://www=
.digitalocean.com/blog/coming-soon-gpu-droplets-nvidia-b300s?utm_medium=3De=
mail&utm_source=3Dnewsletter&utm_campaign=3DAugust-2025&utm_source=3Dmarket=
o&utm_medium=3Demail&utm_campaign=3Dai-newsletter-jan-2nd-edition&utm_conte=
nt=3Db300-announcement)
Featuring 7X more AI compute than NVIDIA Hopper platforms, 2.1TB of HBM3e m=
emory, and high-performance networking integration with NVIDIA ConnectX-8 S=
uperNICs, Blackwell Ultra delivers breakthrough performance on the most com=
plex workloads from agentic systems and reasoning, to real-time video gener=
ation. [Contact](https://www.digitalocean.com/company/contact/sales?utm_med=
ium=3Demail&utm_source=3Dnewsletter&utm_campaign=3DAugust-2025&referrer=3Dp=
roducts/gradient/gpu-droplets&utm_source=3Dmarketo&utm_medium=3Demail&utm_c=
ampaign=3Dai-newsletter-jan-2nd-edition&utm_content=3Db300-contact-sales) o=
ur team to learn more about how to power your workloads with B300.
Follow us:
[Twitter](https://twitter.com/digitalocean?utm_medium=3Demail&utm_source=3D=
newsletter&utm_campaign=3DAugust-2025)
[Facebook](https://www.facebook.com/DigitalOceanCloudHosting?utm_medium=3De=
mail&utm_source=3Dnewsletter&utm_campaign=3DAugust-2025)
[YouTube](https://www.youtube.com/channel/UCaPX53JLxxSbwZz_Ra_cL0g?utm_medi=
um=3Demail&utm_source=3Dnewsletter&utm_campaign=3DAugust-2025)
[Instagram](https://www.instagram.com/thedigitalocean?utm_medium=3Demail&ut=
m_source=3Dnewsletter&utm_campaign=3DAugust-2025)
[Linkedin](https://linkedin.com/company/digitalocean?utm_medium=3Demail&utm=
_source=3Dnewsletter&utm_campaign=3DAugust-2025)
[Twitch](https://www.twitch.tv/digitaloceantv?utm_medium=3Demail&utm_source=
=3Dnewsletter&utm_campaign=3DAugust-2025)
Give $200, get $25. [Refer a friend](https://cloud.digitalocean.com/account=
/referrals?utm_medium=3Demail&utm_source=3Dnewsletter&utm_campaign=3DAugust=
-2025&utm_medium=3Demail&utm_source=3Dfooter&utm_campaign=3Dreferral).
Trouble seeing this email? [View it as a web page](https://anchor.digitaloc=
ean.com/index.php/email/emailWebview?utm_medium=3Demail&utm_source=3Dnewsle=
tter&utm_campaign=3DAugust-2025).
Copyright DigitalOcean. All rights reserved.
105 Edgeview Dr., Suite 425, Broomfield, CO 80021
This email was sent to hi@josie.lol. You can update your email subscription=
preferences here: https://anchor.digitalocean.com/Email_Preferences.html?m=
kt_unsubscribe=3D1&mkt_tok=3DMTEzLURUTi0yNjYAAAGfmmYcZXBjwiAVkrYNdfZ9EMuJH6=
5QPPLvOfz3yIc4TQj_BMDMQ68QWp6z2V7GhyzWDmxxlNLGOTtniBEBP33sKBIlHEYfq4NnCIfcU=
kxkUmdI.
------=_Part_230964148_2081086493.1769449811056
Content-Type: text/html; charset=UTF-8
Content-Transfer-Encoding: quoted-printable
[ DigitalOcean logo. This image is linked to the DigitalOcean homepage. ] [=
[https://www.digitalocean.com/?utm_medium=3Demail&utm_source=3Dnewslett=
er&utm_campaign=3DAugust-2025&utm_medium=3Demail&utm_source=3Dn=
ewsletter&utm_campaign=3DJune-2025]]
ICYMI AI: The latest AI trends, delivered t=
o your inbox
Trend watch: Shadow AI
As AI adoption in the workforce grows, more=
AI-based tools are available to curious developers. This raises concerns a=
bout =E2=80=9Cshadow AI=E2=80=9D which is AI tooling that=E2=80=99s in use =
at a company without IT approval. This might look like an engineer download=
ing Google=E2=80=99s Antigravity IDE on their work laptop without waiting f=
or security review, or a teammate using Claude Code from their terminal whi=
le the official tooling request sits in procurement limbo.
Shadow AI poses security threats to organiz=
ations=E2=80=94and expands the overall attack surface for data breaches and=
sensitive information leaks. According to Netskope=E2=80=99s 2026 Cloud and Threat Report, the three categories of data most involved in=
AI data policy violations over the past year were source code (42%), regul=
ated data (32%), and intellectual property (16%). For example, Samsung discovered that its proprietary semiconductor data was leaked =
through internal engineering use of ChatGPT to check source code, optimize =
code, and record meeting notes for a presentation.
Without the right guardrails and oversight,=
shadow AI will continue to outgrow organizational tooling and systems, fur=
ther putting data at risk. Whether you manage AI tools within an organizati=
on or use them yourself, addressing the risk of shadow AI keeps business and customer data safe.
How to keep shadow AI at bay
For developers, be an advocate, not a rogue=
agent. Test new tools on personal devices, document the productivity gains=
, and bring a case to your team lead or IT. You're more likely to get appro=
val when you show up with findings instead of a request.
For leaders and IT administrators, rigid si=
x-month procurement cycles are part of the problem. Consider faster evaluat=
ion pipelines or sandboxed trials, processes that let teams experiment safe=
ly without pushing them toward workarounds, so your team can stay competiti=
ve without compromising security.
AI trend reports and indus=
try commentary
Ads have landed in ChatGPT: OpenAI announced that it would start testing ads on the Free and Go ti=
ers, noting that a full-scale rollout is not yet planned.
A joint partnership between Google and Apple: Google and Apple have announced a multi-year collaboration to build Ap=
ple Foundation Models on Google's technology, reducing OpenAI=E2=80=99s com=
petitive hold in the market.
Models on our radar
GLM-Image: GLM-Image is an image generation model that adopts=
a hybrid autoregressive + diffusion decoder architecture. In general image=
generation quality, GLM=E2=80=91Image aligns with mainstream latent diffus=
ion approaches, but it shows significant advantages in text-rendering and k=
nowledge=E2=80=91intensive generation scenarios.
Pocket-TTS: A lightweight text-to-speech (TTS) application de=
signed to run efficiently on CPUs. Forget about the hassle of using GPUs an=
d web APIs serving TTS models. With Kyutai's Pocket TTS, generating audio i=
s just a pip install and a function call away.
Flux.2 Klein: The FLUX.2 [klein] model family are our=
fastest image models to date. FLUX.2 [klein] unifies generation and editin=
g in a single compact architecture, delivering state-of-the-art quality wit=
h end-to-end inference in as low as under a second. Built for applications =
that require real-time image generation without sacrificing quality.
GLM-4.7: GLM-4.7, labeled "your new coding partner," =
delivers major improvements over GLM-4.6 across core coding, reasoning, UI =
generation, and tool use, with strong gains in multilingual and agentic cod=
ing and terminal-based tasks=E2=80=94scoring 73.8% (+5.8%) on SWE-bench, 66=
.7% (+12.9%) on SWE-bench Multilingual, and 41% (+16.5%) on Terminal Bench =
2.0=E2=80=94while also supporting "thinking before acting" for better perfo=
rmance in popular agent frameworks like Claude Code, Kilo Code, Cline, and =
Roo Code.
Agent-CPM Explore: AgentCPM-Explore is an agent found=
ation model with only 4B parameters in its entire training and inference in=
frastructure. AgentCPM-Explore has successfully entered 8 classic long-hori=
zon agent benchmarks, including GAIA, HLE, and BrowserComp. AgentCPM-Explor=
e achieves SOTA performance at the same parameter scale and demonstrates it=
s accurate deep research capabilities, effectively breaking the performance=
bottleneck for on-device agents.
Com=
munity tutorials and resources
Agent skills are folders of instructions, scripts, and resou=
rces that an LLM can load to perform tasks. First introduced by Claude in 2=
025, this article creates an Agent Skill for a PDF parser.
LangSmith is a platform for instrumenting, debugging, evaluating, and =
monitoring LLM apps to overcome these challenges and provide end-to-end vis=
ibility into your agent's behavior. In this tutorial, we will go through a =
typical agent reliability loop: instrumenting an agent with tracing, debugg=
ing agent runs automatically and manually, evaluating its performance, iter=
ating on prompts, and finally, monitoring the agent in production.
A t=
rue step forward for open-source video generation, LTX-2 is undoubtedly the=
most powerful video generation technology ever open-sourced. In this tutor=
ial, we will walk through how LTX-2 works by exploring its architectural pipeline, then show how=
to run LTX-2 on DigitalOcean Gradient using the ComfyUI.
Ver=
cel's slick Git-based deploys and edge functions have made it a frontend fa=
vorite, but it's not the only option. Compare Vercel alternatives, including DigitalOcean App Platform, with breakdo=
wns of CI/CD workflows, scaling, and pricing.
Nee=
d to store petabytes of unstructured data without blowing your budget on eg=
ress fees? Check out these object storage solutions=E2=80=94we compare S3 compatibility, pricing =
models, zero-egress options, and AI/ML pipeline integration across provider=
s from hyperscalers to lean alternatives.
Cho=
osing a VPS provider shapes everything from deployment speed to long-term c=
osts. DigitalOcean vs AWS Lightsail breaks down pricing transparency, egress=
fees, performance consistency, and support tiers to help you decide which =
platform fits your workload.
Video tutorials
How AI models run on GPUs (CUDA vs. ROCm) Ever won=
dered how your Python code actually runs on a GPU? This video breaks down t=
he AI compute stack=E2=80=94the layers of technology that transform your hi=
gh-level PyTorch commands into calculations running on thousands of GPU cor=
es. Whether you've heard of NVIDIA's CUDA or AMD's ROCm but never understoo=
d what they actually do, this explainer will give you the complete picture =
from silicon to Python.
Product updates
GPU Droplets accelerated by NVIDIA HGXTM B300 are coming soon Featuring 7X more AI compute than NVIDIA Hopper platforms, 2.1T=
B of HBM3e memory, and high-performance networking integration with NVIDIA =
ConnectX-8 SuperNICs, Blackwell Ultra delivers breakthrough performance on =
the most complex workloads from agentic systems and reasoning, to real-time=
video generation. Contact our team to learn more about how to power your workloads with =
B300.