Durham

I like many things, most of all my ability to like many things.
Nanotech Research Scientist at UC Berkeley.

[Durham @DurhamVSmith]

Anyone have recommendations for payment providers in South Africa that they have actually used.

I have used a few, most recently @paystack but their @PaystackSupport has been horrible with replying to my messages/ hasn't done anything after 25 days and multiple emails.

[Durham @DurhamVSmith]

The coolest AI side project I've seen this year was built with a CNC machine, a USB microscope, and Claude.

It's called AutoProber. A hardware hacker named Jon Gaines shipped it last week.

What it does: turns a cheap hobby CNC into an AI-driven flying probe station. The kind of machine that normally costs R200k+ from commercial vendors, used to reverse engineer circuit boards pin by pin.

The workflow is wild:

💻 You point it at a new board on the plate
👁️ A USB microscope takes frames while the CNC moves around
🧠 Claude stitches the images, identifies pins, pads, and chip labels
✅ Proposes probe targets in a web dashboard for you to approve
⚡ Probes the approved pins and reports the signals back

An oscilloscope channel acts as a safety endstop so the agent can't crash the probe into the board. That detail matters. AI is probabilistic. Hardware is not.

Total parts bill for someone with a 3D printer: a few thousand rand.

For builders in South Africa 🇿🇦, the gap between R200k and R5k is the whole story. Commercial flying probes were never going to land in your workshop. This one could.

Here's the part I keep coming back to:

Most people building with agents try to automate the whole loop. See, decide, act. This does the opposite. The agent narrows the problem from "a thousand possible pins" to "here are twelve worth probing." The human picks. The machine executes. You end up with a system where the agent's job is to make the human's decision easier, not to replace it.

Commercial flying probes will still win in a factory. But the ceiling for what one curious person with a CNC and an agent can build just jumped. That's the interesting part. Not the machine. The precedent.

[Durham @DurhamVSmith]

@spirilis @RaffaySajjad @bcherny Unfortunately this didn't work for me. If its not fixed by the weekend I'll dig a bit deeper into it as well and try push a robust fix.

[Durham @DurhamVSmith]

@spirilis @RaffaySajjad @bcherny I've checked all of these things and like I stated originally getting model access is not the problem (API and playground works fine). It's a Claude Code only bug.

[Durham @DurhamVSmith]

@TonyMcDow @jun_song Measuring things is negativity now?

[Durham @DurhamVSmith]

@TimothyBrown @bcherny The writeup and a fix I wouldn’t trust is in the GitHub

[Durham @DurhamVSmith]

@jeetganatra1 @bcherny IIRC you still need explicit requests for anthropic’s models (and maybe some vendors I don’t use)

[Durham @DurhamVSmith]

@bcherny I do. I can access it via api and model playground no problem. The update from 4.5 to 4.6 was smooth as butter but this one has had some hairs.

[Durham @DurhamVSmith]

@ClaudeDevs

[Durham @DurhamVSmith]

Is anyone else having issues trying to access @AnthropicAI Opus 4.7 via @claudeai Code with the actual provider being @awscloud Bedrock. It's not working but accessing 4.7 via API or model playground is working fine.

[Durham @DurhamVSmith]

🇿🇦 SA business owners using Fiverr: every file you've ever sent through their chat might be public right now.

Not "theoretically accessible." Indexed by Google. Searchable by anyone. Right now.

Here's what happened: a security researcher found that Fiverr stores every file shared between client and freelancer on public URLs.

The line of code that blocks Google from crawling these files? It's in their robots.txt. Commented out. Sitting there doing nothing. 😬

The researcher reported it to Fiverr's security team 40 days ago. No reply. So it went public.

What people are already finding in Google search results:
1️⃣ Tax returns with full personal details. Thousands of them
2️⃣ Admin passwords to entire websites, including ones people could be blackmailed with
3️⃣ API tokens and credentials that have probably never been changed
4️⃣ Confidential business documents, NDAs, internal APIs
5️⃣ Penetration test reports (yes, security reports, publicly searchable)
6️⃣ Paid digital products and courses. Free for anyone to download

The kicker? Fiverr holds an ISO 27001 security certification. And they actively buy Google Ads for tax filing keywords while the resulting work product sits on a public URL.

Why this matters in SA:

Fiverr is the first stop for a lot of SA businesses. Quick logo. WordPress site. Someone to do your books or file your returns. That freelancer who set up your email and hosting.

Think about what you sent through that chat. CIPC documents. Banking details. Login credentials you asked the freelancer to use. Client files.

All of it is sitting on a public URL that Google has already indexed.

Under POPIA, if your customer's personal information leaked because you sent it through an unsecured platform, that's on you. Not Fiverr. You. ⚠️

What to do today:
1. Search "site:https://t.co/a38fauI8nw" plus your business name or any document you remember sending
2. If you find anything, screenshot it and request removal through Google's search console
3. Change every password you've ever shared through Fiverr chat. Today
4. Stop sending sensitive files through platform messaging. Use a secure file share

Fiverr had 40 days to fix this. They didn't even reply.

Don't give them another one.

Not sure where to start? DM me. I'll help you check if your business info is exposed. 🇿🇦

[Durham @DurhamVSmith]

Something I'm seeing from the inside, and it's starting to worry me.

I run a web agency in South Africa🇿🇦. Hundreds of local businesses pass through our doors every month. Plumbers, lawyers, dentists, logistics companies, retailers.

Almost all of them run their entire business on R49/month email and hosting from a two-man shop that hasn't updated security in years.

That's where their contracts live. Client IDs. Invoices. Salaries. Every email they'd never want read out loud.

This was already a problem. It's about to become a disaster.

This week, Anthropic pulled together Amazon, Apple, Google, Microsoft, Cisco, NVIDIA, Broadcom, CrowdStrike, Palo Alto Networks, JPMorgan Chase, and the Linux Foundation. Eleven of the biggest names on earth. For one reason:

They built an AI so good at finding security holes, they needed each other to help contain it.

In a few weeks of testing, it found thousands of previously unknown security holes in every major operating system and browser on the planet. One had been hiding inside OpenBSD, one of the most secure systems ever built, for 27 years. Missed by every human expert. Found by AI in days.

The smartest companies on earth are scrambling. Because the moment this capability leaks to the bad guys, anything weak gets cracked open in minutes, for almost free.

Now look at the average SA business.

The global giants are arming up with frontier AI. The local budget host is running software from 2006 on a server in someone's spare room.

That gap will get exposed. And it won't be the banks making headlines. It'll be ordinary SA businesses waking up to find every document they've ever sent posted on a forum.

Nobody local is talking about this. So I will.

If you're unsure whether your host is safe, DM me. I'll give you a straight answer. No pitch.

[Durham @DurhamVSmith]

https://t.co/sQavhog46x

[Durham @DurhamVSmith]

@Alibaba_Qwen just open sourced Qwen3-TTS and it's genuinely impressive voice cloning from a 3 second sample, voice design from text descriptions, and high-quality preset speakers.

I've been playing around with @ManimCommunity a lot recently for some educational content and wanted a way to add natural-sounding voiceovers without leaving Python. So I built a plugin that integrates Qwen3-TTS with manim-voiceover.

If you've used manim-voiceover before, the interface is familiar.

There are three services depending on what you need:
- Qwen3VoiceCloningService: clone any voice from a short audio sample
- Qwen3VoiceDesignService: describe a voice in plain English and it generates one
- Qwen3PresetVoiceService: 9 built-in voices with emotion/style control via an instruct parameter

The video shows all three in action — the narrator is cloned from everyone's favorite big uncle @SnoopDogg , the hero uses a preset speaker, and the mentor/villain are both designed purely from text
descriptions.

Runs locally on ~4GB VRAM. 10 languages supported.

Link to the repo is in the comments.

[Durham @DurhamVSmith]

Breakfast in bed 🥜🐿️

[Durham @DurhamVSmith]

@irfanrosandi @kieranklaassen @AnthropicAI @mikeyk @_catwu I didn’t see this, but the reason, I think, is that it is used for context window compacting.

[Durham @DurhamVSmith]

Life took an unexpected turn on the evening of December 18th. I found this little one in my front garden, all alone. Sadly, we had found her sibling who didn’t make it earlier that morning. We waited for hours, hoping the mother would return, but she never came.
When I called the rehab centers, the news wasn’t good—they said their only option would be to euthanize her. My heart just couldn’t let that happen. So, looks like I’m officially a foster father for the next few months. She’s safe, warm, and sleeping soundly in my hand. 🐿️❤️

[Durham @DurhamVSmith]

@michael_nielsen One could be try something like a council of LLMs, each with different system prompt personalities and temperature settings to identify possible sources of misinterpretation and help refine the question. Worth the effort? Maybe not but an interesting experiment.