Madison May (@pragmaticml) header
Madison May (@pragmaticml) profile picture

teaching machines @indicodata - professional novice

Asheville, NC
Joined March 2010
Photo by Madison May (@pragmaticml)Photo by Madison May (@pragmaticml)Photo by Madison May (@pragmaticml)Photo by Madison May (@pragmaticml)Photo by Madison May (@pragmaticml)Photo by Madison May (@pragmaticml)Photo by Madison May (@pragmaticml)Photo by Madison May (@pragmaticml)Photo by Madison May (@pragmaticml)Photo by Madison May (@pragmaticml)
This week's blog post explores methods for incorporating longer-term context in transformers! Featuring 6 unique approaches: - Sparse Transformers - Adaptive Span Transformers - Transformer-XL - Compressive Transformers - Reformer - Routing Transformer pragmatic.ml/a-survey-of-met…
7
70
4
294
Madison May retweeted
I finally caught one in the wild. too funny.
18
5
2
507
118,849
Madison May retweeted
🚨 Microsoft’s agent-framework package on PyPI was compromised. Malicious versions 1.26.1 and 1.26.2 use Python startup hooks to steal Azure, GitHub Actions, and other CI/CD credentials. Technical analysis + IOCs → socket.dev/blog/agent-framew…
15
44
15
189
22,131
Interesting tension -- agents are the clearest fit for software that has a "compact description" (e.g. is a "deep module") and a small "blast radius". - Core infra: compact description but large blast radius - Application layer: verbose description but small blast radius
123
Good software is "bigger on the inside" (i.e. is composed of deep modules). Software with this property tends to be easier to verify. Verifiable software is agent-friendly software.
1
2
168
Was so excited to introduce my non-technical friend to GitHub yesterday and expound on the merits of version control, only to have GitHub proceed to error out on every other operation. Me: "Oh yeah that's normal, don't worry about it, just try again"
1
230
Pin your GitHub actions workflows to specific hashes
🚨 Massive GitHub Actions CI Secret Stealer 4,800+ compromised accounts attempted to push malicious workflows to 61k+ repos. Payload dumps all CI secrets, memory, OIDC & commit secrets, then instantly deletes run logs. Search your repos for: • PR: "docs: dependency audit notes" • Branch: wfx-docs or chore/ • Commits: "ci: refresh build cache" or "chore: add ci helper" • Actions tab: Skipped/missing run numbers If hit, assume total compromise. Rotate ALL secrets and cloud roles immediately.
2
1
171
Madison May retweeted
Engineers at @SocketSecurity do what's right for the business, even when it's unglamorous. Our mantra is "you can just ship things." We have no PMs, so engineers work directly with customers. And we're winning. We're hiring engineers. socket.dev/careers
2
3
3
18
2,817
All API keys should come with a special "burn" permission that allows disabling the key. That way, whitehats can immediately disable any keys they find in the wild. (credit to the ever-brilliant @seveibar for this idea!)
54
69
7
1,827
89,932
Madison May retweeted
Today we are launching Raindrop for Good. Raindrop is free for eligible nonprofit and public benefit projects building agents. Apply here: raindrop.ai/for-good/
3
11
3
38
2,966
Security features should not require enterprise plans. Use tiered pricing for higher usage or special features, don't make startups trade off security and budget. Your own security posture might benefit too!
i hate how many security features are gated behind expensive enterprise plans so many products are impossible to use safely even for a small team with agents hacking everything this feels like a bad place to upsell your users
1
204
UMAP 0.6 is out! A new initialization, new optimizers and faster nearest neighbour search make UMAP much faster and far better at large scale. On 3 million arXiv papers it finishes in under 5 minutes, before UMAP 0.5 has even found its nearest neighbours.
24
110
18
604
26,717
Replying to @a_musingcat
slop means it takes less human effort to produce than it does to consume (??)
1
2
10
814
Madison May retweeted
This is so on-point. We are now past the point where the basic model is interesting and well into the age of nuance of how the system around the model works
3
13
1
157
4,851
I respect this kind of absolute transparency. Give your users tools to hold you accountable so that your incentives align with theirs
other actors in the sandbox space often try to attack us on stability and speed but we're the only ones transparent enough to publicly show you uptime feature by feature, hour by hour AND fluctuation of key performance metrics / p99s over time contrary to some bigger providers with plenty outages and nothing on their status page about them... if you use them you know status page linked in our landing page or over there (will update domain next thing I promise): status.ascii.dev
1
124
"Software factory" is a product category that will be won by great designers rather than great engineers
82