Neil Ferguson

567 posts

Neil Ferguson

@nfergu

Views are my own and do not represent the views of my employer

Cambridge, UK Katılım Şubat 2009

1K Takip Edilen121 Takipçiler

Neil Ferguson retweetledi

Feargal Sharkey@Feargal_Sharkey·2 Eyl

Almost 8,000 potentially illegal spills happened on dry days during the first half of the year and South West Water was responsible for more than a quarter." So how's that getting tough on the water industry thing working out I wonder? thetimes.com/uk/environment…

English

587

1.3K

17K

Neil Ferguson retweetledi

Carole Cadwalladr@carolecadwalla·11 Haz

More journalists killed in Gaza than World War 1, World War 2, Vietnam, Yugoslavia & Afghanistan combined

Barry Malone@malonebarry

Al Jazeera has published the names of every journalist killed since Oct 7th. 231 Palestinian journalists have been slaughtered in Gaza. A whole generation of reporters wiped out while many of their colleagues in the West simply shrugged.

English

1.2K

11.7K

24.8K

1.1M

Neil Ferguson@nfergu·29 Oca

@StephenPiment At 6 tokens per second. To run it at a reasonable speed is at least $500k. Plus power costs. Or you could pay ~$200 per hour to run it in the cloud.

English

209

Stephen Pimentel@StephenPiment·29 Oca

You can quarrel all you like about how much it cost DeepSeek to train R1, but we know how much it costs to host and run: $6000.

Matthew Carrigan@carrigmat

Complete hardware + software setup for running Deepseek-R1 locally. The actual model, no distillations, and Q8 quantization for full quality. Total cost, $6,000. All download and part links below:

English

448

205.1K

Neil Ferguson@nfergu·29 Oca

@pmarca At 6 tokens per second, with a small context window. Seems like a bad deal to me. Might save on heating though.

English

Marc Andreessen 🇺🇸@pmarca·29 Oca

$6,000 to run full Deepseek R1 LLM with reasoning at home. 🤖🚀

Matthew Carrigan@carrigmat

Complete hardware + software setup for running Deepseek-R1 locally. The actual model, no distillations, and Q8 quantization for full quality. Total cost, $6,000. All download and part links below:

English

422

666

6.8K

1.1M

Neil Ferguson@nfergu·28 Oca

@azeem Reference for Philip Schmid’s post: linkedin.com/posts/philipp-…

English

Neil Ferguson@nfergu·28 Oca

@azeem Thanks! I think the analysis is wrong. I’m not an expert on MOE models, but AFAIK they choose a subset of the parameters very often (every token?) during inference. So they save on compute but not on memory requirements. Phillip Schmid says 16xH100s, which seems correct to me.

English

Neil Ferguson@nfergu·28 Oca

@azeem you say in today’s post that Deepseek can “run on just two consumer-grade Nvidia 4090 GPUs (costing roughly $2,000)”. Are you sure about this? It doesn’t sound right to me.

English

Neil Ferguson retweetledi

Carole Cadwalladr@carolecadwalla·6 Eki

Journalists are meant to speak truth to power. That's our job. But what do you do when that power is your own news org? This thread (professional suicide note?) is about @guardian & @ObserverUK's future. Because it turns out you can’t believe everything you read on a poster.. 1/

English

212

2.9K

8.2K

1.8M

Neil Ferguson@nfergu·24 Eyl

@justcookit Wondering if you’ve tried Rishright’s Yo Po Chilli? Asking because we had a conversation about a different brand of chilli condiment years ago, and Rishright is a Cambridge-based company and IMO superior to the other brand.

English

Neil Ferguson@nfergu·21 Tem

@cstanley Tony Hoare called it his billion dollar mistake: infoq.com/presentations/… That’s an underestimate on this case I think.

English

Neil Ferguson@nfergu·4 Tem

@Cloudflare Seems like a nice feature but I’m skeptical about this claim. How would you know if you got a false negative if you don’t know the ground truth?

English

428

Cloudflare@Cloudflare·3 Tem

To help preserve a safe Internet for content creators, we’ve just launched a brand new “easy button” to block all AI bots. It’s available for all customers, including those on our free tier. Read our blog post for more details: cfl.re/3RQYlxz

English

323

1.9K

337.3K

Neil Ferguson@nfergu·8 May

@raymondh Thanks! I probably don’t have any way to avoid this in my case, so I’m catching the ValueError at the moment.

English

Raymond Hettinger@raymondh·7 May

@nfergu Here, a call to r() will cause get's g.__closure__[0].cell_contents to raise a ValueError. def f(): x = 10 def get(): return x def remove(): nonlocal x del x return get, remove g, r = f()

English

398

Raymond Hettinger@raymondh·7 May

#Python tip: When debugging, testing, or explaining closures, your new best friend is: >>> pprint.pp(inspect.getclosurevars(func)[0]) It is like locals() but for closure variables.

English

139

13.4K

Neil Ferguson@nfergu·29 Ara

@BrianRoemmele @stevesi Are you going to tell us how you did this?

English

Brian Roemmele@BrianRoemmele·28 Ara

I have done extensive work studying The Pile— an open source dataset for training AI foundational models. I have been able to “replicate” “news stories” that this dataset NEVER saw—almost verbatim. I have a theory how and why this happens and it changes any lawsuit and claims.

English

209

1.2K

607.1K

Neil Ferguson retweetledi

Jason Kint@jason_kint·28 Ara

ok, I've now read the full NYT complaint filed this morning vs OpenAI and Microsoft. I'm impressed - it's future-focused around fair value for work vital to democracy. It also contains 220k pages of exhibits although the pages of Ex J stood out to me. more on that in a minute. /1

English

206

2.2K

10.8K

3.6M

Neil Ferguson@nfergu·4 Ara

@paulg Lots of brilliant young adult books. He should read the Cherub series by Robert Muchamore.

English

197

Paul Graham@paulg·4 Ara

Taught 11 yo about the concept of "young adult fiction" and that he should just skip it and go straight to books written for adults.

English

192

2.1K

Neil Ferguson@nfergu·6 Kas

@PaulWallace123 Think you mean @nfergus. I know almost nothing about inflation, or wars (I do have have GCSE in economics, but all I can remember is something about mars bars and marginal utility).

English

Neil Ferguson@nfergu·11 Eki

@SparxLearning is there any way in Sparx Maths that my daughter can see the questions she has answered for completed assignments? I’d find this really useful for reviewing assignments with her, to see if there was anything she struggled with etc

English

Neil Ferguson@nfergu·2 Eki

@iamramiawar @GergelyOrosz What serverless GPU provider are you using?

English

Rami Awar@iamramiawar·2 Eki

@GergelyOrosz I started an ML team at my company to extract different details from uploaded documents. So far, hosting 4 different fine-tuned models on a serverless GPU provider that also provides API interfaces to the models. Fine tuning was done offline.

English

3.8K

Gergely Orosz@GergelyOrosz·2 Eki

If you’re a software engineer and your company is integrating AI functionality into products: what is the use case and how are you doing it? Eg using an API for a third-party, self-hosting a model, fine-tuning a model etc?

English

160

94.9K

Keşfet

@StephenPiment @pmarca @azeem @guardian @ObserverUK @justcookit @cstanley @Cloudflare