/void !

DeepSeek Flash 4.1: the open source model putting pressure on the American players

2026-10-13T09:00:00+02:00

Let me be blunt: DeepSeek’s latest model, Flash 4.1, kind of blew my mind.

We’re used to DeepSeek shipping solid models at ridiculous prices. But this time, they went one step further, and not a small one.

Open source, for real🔗

The first thing is that Flash 4.1 is open source. Not “half open weights”, not “open but with a license that forbids you from using it for work”. No: the weights are there, you can download them, host them yourself, fine-tune them, run them on your own infra.

That’s a huge difference. When you depend on a closed API, you depend on someone else: their prices, their terms, their outages, their dumb model changes that break your production overnight. Here, you stay in control.

And the price…🔗

That’s the second point, and probably the most annoying one for the competition.

Flash 4.1 costs a fraction of what the American models charge, for a level of quality that, honestly, has little left to envy them for on most everyday tasks. Generating code, writing, summarizing, extracting data, holding a conversation: for 90% of real use cases, the difference has become hard to justify.

When an open, nearly free model does the same job as a closed model billed ten times more, at some point you need a real reason to pay. “It’s a US model” is no longer one.

Real competition for the American models🔗

That’s where it gets interesting. The big American players (OpenAI, Anthropic, Google) built their business on two pillars: performance, and being the only ones able to deliver it.

The first pillar still holds, but less and less. The second one is collapsing.

Because DeepSeek’s strategy is brutal: they publish openly, they slash prices, and they force everyone to go back to the drawing board. Other labs now have to either lower their rates or justify added value that no longer exists for most customers. That’s enormous pressure, and it comes from a player that, only a few years ago, wasn’t even in the race.

What it changes for us🔗

Concretely, for someone like me who tinkers with projects, writes scripts, and uses AI to rough out work: it means we can stop counting every token as if it were gold.

We can test, make mistakes, iterate, and re-run generations without it hurting our wallet. We can also host the model ourselves if we have sensitive data and don’t want to send it off to a third-party provider. In short, we get control back.

Fair warning, though🔗

I’m not going to fall into fanboyism either. An open source model means you handle the infra, the updates, the security. It’s not magic. And on the most demanding tasks, the American frontier models still keep a lead.

But for everything else, the question is no longer “is open source good enough?”. The question has become “why would I pay full price for this?”.

In short🔗

Flash 4.1 isn’t just a good model. It’s a signal: quality is becoming commoditized, prices are collapsing, and the closed American models are going to have to learn to live with competition they no longer control.

And honestly, that’s good news for everyone. Except for their margins.

From « Code Monkey » to Systemic Orchestrator: how the developer job is changing

2026-09-19T15:30:00+02:00

For a long time, a developer’s daily life often came down to one thing: turning an idea into lines of code by hand, loop by loop, semicolon by semicolon.

Today, any LLM can generate a function or a UI in three seconds. Typing code by the mile no longer has any value. The job isn’t dead, but its focus is shifting radically.

Less syntax, more judgment🔗

Generating code has become free. Maintaining and understanding it remains costly. Day to day, our time no longer goes into hunting for syntax or boilerplate code, but into three essential topics:

Expressing the need clearly: If the instruction is vague, the AI produces a bad solution in a flash. Precision in modeling and business rules matters more than ever.
Reviewing with a critical eye: Generated code often looks clean at first glance. That’s misleading: edge cases, security holes, or logical nonsense hide easily inside it. Our main role has become that of a ruthless reviewer.
Hardening the tests: To trust code you didn’t type yourself, you need solid tests (unit, integration, strict typing). The test harness becomes our safety net.

The big picture rather than the isolated brick🔗

AI is good at coding one specific function, but it has a short memory and doesn’t see the big picture. Architecture choices, module breakdown, infra cost management, and database consistency remain purely human decisions.

If you let agents stack code without any control, you quickly end up with an unmanageable codebase that nobody masters: tech debt on autopilot.

What about beginners?🔗

That’s probably the real pain point. You used to become senior by breaking things and spending nights hunting for a bug in your own code.

For a junior today, the trap is using AI as a crutch without ever understanding what runs under the hood. The key to learning is shifting: it’s less about writing from a blank page than about learning to read, debug, and challenge what gets generated.

In short🔗

The job doesn’t disappear, it sheds its most mechanical part. Less time wasted on basic plumbing, more energy invested in design, robustness, and value for the user.

Alacritty: my everyday terminal ⚡

2026-09-19T15:00:00+02:00

I do more and more things from the terminal. For that everyday use, I needed an ultra-efficient tool: Alacritty.

Day to day, Alacritty has become indispensable. Written in Rust and GPU-accelerated, it goes straight to the point.

What wins me over:

🛡️ Robustness: Exemplary stability, it never crashes even under heavy load.
⚡ Speed: GPU rendering means instant keystrokes.
⚙️ Simplicity: A minimalistic, clear configuration file.

A must-have when you live in the terminal!

Herdr: the terminal multiplexer built for AI agents 🤖

2026-09-19T14:30:00+02:00

With the rise of AI coding agents (Claude Code, GitHub Copilot CLI, Cursor Agent…), the way we work in the terminal is changing fast. Running several agents in parallel inside a classic multiplexer (like tmux or zellij) quickly becomes messy when you need to track who’s doing what. That’s where Herdr comes in.

What is Herdr?🔗

Herdr is a persistent terminal multiplexer built specifically for the age of AI agents. It runs as a background server and adds an agent-aware management layer on top of your usual PTYs.

What Herdr lets you do🔗

🚦 Real-time status tracking: Herdr automatically detects each agent’s state and displays it clearly (running, waiting for input, blocked, or done). No more endless scrolling to figure out whether an agent needs you!
🔄 Absolute persistence: Your tasks keep running even if you close your terminal window or disconnect an SSH session.
🔌 API & automation: Through its CLI and socket API, agents can themselves interact with the multiplexer (open new tabs, launch sub-tasks).
🖥️ Native terminal: It integrates seamlessly with your favorite terminal emulator using standard PTYs.

In short🔗

If you use AI agents in the CLI every day for your projects and want to maximize your productivity without losing track, Herdr is a tool you should try right away.

Criteo — SRE

2022-01-01T00:00:00+00:00

I joined the Criteo Webapps scope as an SRE. Keeping a large-scale platform reliable is a daily job, and most of it happens before anything breaks. Between two incidents, I write code, migrate systems and try to make deployments boring.

What I did🔗

Keep the Criteo Client API available and healthy
Write and maintain services in C#, Go and Python
Take part in the Flink migration
Build continuous deployment so changes ship safely

Tech stack🔗

git, gerrit, chef, mesos, jenkins
grafana, kibana, ELK, opsgenie
Python (Flask, Django), Go, C#, Java, Scala
Couchbase, Memcached, Mssql, MariaDB
Hadoop, Hive, Kafka, protobuf, parquet
kerberos, ldap, slack, jira, confluence

Rust app in terminal 🖥️

2021-07-08T09:19:42+01:00

Rust is awesome.

Recently i found some great application for the terminal.

`fd`🔗

find alternative

`exa`🔗

ls alternative

`dust`🔗

bored of du ?

`alacritty`🔗

quickTerm

`spotify-tui`🔗

Spotify in the terminal

🚧 Gitlab and Github pages handling the same code 🤓

2021-04-11T09:19:42+01:00

Hello,

A small trick to have your static website hosted on github + gitlab, with the same source code.

Why ? If one fall of course ! :D And also because it’s fun, at least i think it’s fun :)

Gitlab - Master🔗

create our static website repo
add web stuff inside the repo
configure CI

create our static website repo🔗

let’s pretend our repo name is upcreid.gitlab.io and it’s a hugo static website. Let’s create it and make it public

then the usual stuff:

git clone git@gitlab.com:upcreid/upcreid.gitlab.io.git
cd upcreid.gitlab.io
touch README.md
git add README.md
git commit -m "add README"
git push -u origin master

create gh-pages branch🔗

before going further we’ll need to create a gh-pages branch on gitlab side, this in order to not break the CI when pushing without a gh-pages branch.

Go then !

git checkout --orphan gh-pages
git rm --cached -r .
git commit -m "new gh-pages branch" --allow-empty
git push origin gh-pages

go back to master

git checkout master

add web stuff inside the repo🔗

Ok, now let’s add some code inside to have some static shit in there.

Install hugo
Create new site

hugo new site --force .

Install a theme

git submodule add https://github.com/theNewDynamic/gohugo-theme-ananke.git themes/ananke

add the theme, name and remove baseURL to the conf file

echo $'languageCode = \"en-us\"\ntitle = \"upcreid.gitlab.io\"\ntheme = \"ananke\"' > config.toml

Start the server to verify if everythin goes ok

hugo server

Commit

git add .
git commit -m "add basic hugo"
git push

configure CI🔗

Complicated !

Just add one more file to the repo…

.gitlab-ci.yml

image: registry.gitlab.com/pages/hugo:latest

variables:
  GIT_SUBMODULE_STRATEGY: recursive

pages:
  script:
  - hugo -b "https://upcreid.gitlab.io"
  artifacts:
    paths:
    - public
  only:
  - master

Commit

git add .
git commit -m "add basic hugo"
git push

and voila => https://upcreid.gitlab.io

Now, let’s handle github.

Github - Mirror🔗

create our static website repo
branch configuration
git mirror
configure CI

create our static website repo🔗

Branch configuration🔗

After creation, create a branch called master. To be in phase with gitlab

echo "# upcreid.github.io" >> README.md
git init
git add README.md
git commit -m "first commit"
git branch -M master
git remote add origin git@github.com:upcreid/upcreid.github.io.git
git push -u origin master

Git mirror🔗

Main part of this is to trigger the mirroring on gitlab side. So inside gitlab repo, simply launch

git remote add --mirror=push github git@github.com:upcreid/upcreid.github.io.git

So, we add a remote, and say it’s a mirror (–mirror) and we name it github.

Verify

git remote -v

you should have something like:

github  git@github.com:upcreid/upcreid.github.io.git (fetch)
github  git@github.com:upcreid/upcreid.github.io.git (push)
origin  git@gitlab.com:upcreid/upcreid.gitlab.io.git (fetch)
origin  git@gitlab.com:upcreid/upcreid.gitlab.io.git (push)

configure CI🔗

Then create the github CI file, we’ll use the github actions. But we should commit those github action inside the gitlab repo.

So inside the gitlab repo aka master, create the following dir/file:

mkdir .github/workflows

touch .github/workflows/gh-pages.yml

And inside the gh-pages.yml

name: github pages

on:
  push:
    branches:
      - master

jobs:
  deploy:
    runs-on: ubuntu-18.04
    steps:
      - uses: actions/checkout@v2
        with:
          submodules: true  # Fetch Hugo themes (true OR recursive)
          fetch-depth: 0    # Fetch all history for .GitInfo and .Lastmod

      - name: Setup Hugo
        uses: peaceiris/actions-hugo@v2
        with:
          hugo-version: '0.81.0'

      - name: Build
        run: hugo -b "https://upcreid.github.io"

      - name: Deploy
        uses: peaceiris/actions-gh-pages@v3
        with:
          github_token: ${{ secrets.GITHUB_TOKEN }}
          publish_dir: ./public

And commit

git add .
git commit -m "add github CI"
git push origin && git push github

As you can see, last command is in fact 2 commands. in order to push on gitlab + github. Some tricks can be made here to push with one command => google

And that’s it => https://upcreid.github.io/

Conclusion🔗

Since you got 2 remote you’ll need to push to those remote after commiting.

you can do that simply with

git push origin && git push github

Then, nothing more should be done and both CI will be triggered.

See you !

TheMove 🛴

2021-04-08T09:19:42+01:00

I’ve quit Paris recently, to a more little City.

Because of the Covid stuff, but mainly to be closer to the nature.

It was in my mind since a long time, and the actual crisis, and my work, give me the opportunity to move away.

Remote Work is kind of OK.

I’m missing the office, and mainly the people, but since i’m not alone at home, it’s easier to handle.

I was kind of afraid at the beginning, but i think i found my rhythm.

Git two accounts alias

2019-05-09T00:00:00+00:00

Hello,

Recently, I was faced with a problem. How to manage two git accounts on the same computer? I explain myself, a “personal” account and a “work” account, on the same computer. Yes, I know that working on personal things with work materials is a bad habit … but that’s another question.

So, to get rid of just that (I did not invent anything, there are a lot of tutorials on the web), you will need to create a separate directory. One for work, one for the person.

Regarding git, you will have to manage your \ $ HOME .gitconfig like this:

[includeIf "gitdir:~/perso/"]
    path = ~/perso/.gitconfig

[includeIf "gitdir:~/work/"]
    path = ~/work/.gitconfig

[user]

So as you see, [user] is not define. So you’ll have to define it in another .gitconfig. So in each directory, create a .gitconfig

in the /perso directory:

[user]
   email = perso@perso.com
   name = name

in the /work directory:

[user]
   email = work@work.com
   name = name

And that’s it. Each time you’ll go into the folder, your “git user” will change.

Bonus🔗

Here is my git alias for log:

yes there is 3, 3 differents view.

[alias]
    lg1 = log --pretty=format:"%C(yellow)%h%Cred%d\\ %Creset%s%Cblue\\ [%cn]" --decorate
    lg2 = log --pretty=format:"%C(yellow)%h%Cred%d\\ %Creset%s%Cblue\\ [%cn]" --decorate --numstat
    l = log --graph --pretty=tformat:'%Cred%h%Creset -%C(yellow)%d%Creset %s %Cgreen(%an %cr)%Creset' --abbrev-commit --date=relative

See you !

Criteo — Escalation Engineer

2018-02-11T00:00:00+00:00

Escalation Engineer is the last line before things go really wrong. The job was to stop the flood, protect the business, and keep developers, support and clients talking to each other — and to write code whenever it actually helped.

What I did🔗

Stop the flood and contain incidents
Protect the business under pressure
Handle incidents end to end
Support the development teams
Connect the right people with each other
Write code whenever I could

Interview🔗

I talked about this role in a Medium article.

Tech stack🔗

git, gerrit, chef, mesos, jenkins
grafana, kibana, ELK, opsgenie
Python (Flask, Django), Go, C#, Java, Scala
Couchbase, Memcached, Mssql, MariaDB
Hadoop, Hive, Kafka, protobuf, parquet
kerberos, ldap, slack, jira, confluence

Veolia — Treasury suite

2015-04-11T00:00:00+00:00

I was the technical consultant on a treasury management suite (Wallstreet Systems). Over those two years, the platform moved out of its old hosting, got connected to more banks, and had to keep running the whole time.

What I did🔗

Analysis and delivery🔗

Python tooling to inject bulk transactions from CSV for new entities
BugZilla migration from Windows to Linux (apache2, mysql, Bugzilla)
Accounting exports (xml, xsl) to SAP and SAGE
Support for outsourcing the back office (Next → Datalog): MT940, then MT101 / MT210 / MT300
Reference data export (banks, entities, accounts) to Datalog
FxAll rollout with Reuters and WSS
Hosting migration: HP-UX → Linux, ISO environment (PROD replica at D-1, RMAN), network setup (two datacenters, MPLS), operations review

Continuous improvement🔗

Development, packaging, integration and delivery
Stored procedures to speed up reporting
XML framework for banking flows: AFB120, MT101, MT940
Change control in SVN
SQL tuning (views, indexes, triggers)
Reports and activity dashboards in the WSS suite

Tech stack🔗

HP-UX, Windows Server 2003, Citrix
Oracle, SQL Developer, PL/SQL
Weblogic, WSS Suite (CMM, TRM, ACM), activeMQ
SVN
shell, Python, Perl
SWIFT Alliance, esiadapter, FTP

Caisse des dépôts

2013-09-11T00:00:00+00:00

A data migration project for a public financial institution: take decades of data out of legacy systems and put it into a clean Oracle database, without losing a single row. Half engineer, half translator between the business and the systems.

What I did🔗

Solutions analyst🔗

Shell scripts to recover legacy data (MVS, DB2, Oracle, Sybase, XML)
Clean, enrich and reshape the data before loading
Load it into Oracle (reference data, derivations, banking flows)
Reconcile counts between source and target
Update secure file transfers (CFT)
Batch scripting to format data automatically (WinSCP, Excel)
Orchestrate the jobs with Control-M
Write and tune SQL queries
Package everything in a dedicated SVN repository (KSH, SQL, ENV)
Train the team on SVN and continuous delivery habits
Build shell functions used across the CDC infrastructure
Validate deliverables (technical documents, product)

Business analyst🔗

Run workshops with business teams to define migration rules from source analysis
Analyse the data to migrate
Define the migration sequence and rules
Assist the business teams during acceptance testing
Define flow matrices and configure the transfer tools
Write documentation and communicate in French and English

Tech stack🔗

CFT (SOPRA)
Unix, KSH, Windows Server 2003, VBS, Citrix, Cisco
Oracle, Sybase, MVS, DB2
SVN, Quality Center, JIRA, Control-M

Veolia — Next treasury project

2011-08-11T00:00:00+00:00

Technical consultant on the Next treasury project (Wallstreet Systems, SwiftNet, Reuters). I owned the production environments and helped carry a major version upgrade all the way from integration to go-live, while keeping support and operations in the loop.

What I did🔗

Integration and environments🔗

Integrate every vendor release across the integration platforms
Set up monitoring to see how each delivered version behaves
Shell scripts for filesystem checks, log purging and backups
Daily database refresh for the data warehouse (RMAN, SQL, shell)
Lead the setup of file transfers with other Veolia entities (Dalkia, transport, water…)
Build an analysis database from application logs
SQL reporting and control queries
Manage SVN configuration across INT / REC / VAL / PROD with the VEIT production teams
Create SVN branches to track each environment
Analyse and set up banking flows
Translate user needs into technical specs for the vendor
Write functional specs and operations runbooks

TMA / L2-L3 support🔗

Train L2 and L3 support on the new version
Write the installation procedure
Package deliverables for production
Optimise production (SQL analysis, vendor feedback, Oracle and server config)
Set up alerting and email notifications
Communicate and escalate on incidents
Manage disaster recovery plans
Adopt ITIL practices (change management)

Tech stack🔗

HP-UX, Windows Server 2003, Citrix, IIS
Oracle 10g, SQL, PL/SQL, SQL Developer
Weblogic, Apache, WSS Suite, activeMQ
Java, shell, Python, Perl
SWIFT Alliance, esiadapter, RMDS, CFT, FTP, Samba
SVN, Centreon, ServiceNow

Accor

2007-09-11T00:00:00+00:00

Production analyst in a busy operations cell. My days were about watching, fixing and shipping — making sure applications stayed up, batches ran on time, and the marketing team got its data.

What I did🔗

Supervise production and partner interfaces
Write operations procedures
Test, validate and deploy applications every week
Watch, fix and launch daily batches
Configure and run Nagios
Pull SQL extracts from customer databases for marketing
Write and adapt SQL correction scripts
Transform and integrate flat files
Support users and the dev teams during releases
Onboard new members of the team

Tech stack🔗

Java, Struts, Hibernate, Eclipse, Maven, Archiva, Continuum
Debian, Red Hat, Tomcat, Apache, Resin, Joomla
Informix, Postgres, Talend, SQuirreL, pgAdmin
shell, XSD, XSL
CFT, SSH, GPG, SFTP, Crontab, Quartz, Squid

Opéra National de Paris

2005-01-11T00:00:00+00:00

Operations technician at the Paris Opera. A small team, critical systems, and a lot of care: the shows had to go on, and the applications behind them had to keep running.

What I did🔗

Update, query and run recurring jobs on the Opera database
Check that applications were running on the servers
Manage production backups
Write operations procedures
Configure the PLEIADES package (requests, EDI, etc.)
Write and change SQL queries through Business Objects
Develop small Visual Basic applications
Support users on an Excel planning macro used by the Opera managers
Build mail-merge scripts

Tech stack🔗

Business Objects, PLEIADES
Visual Basic, Excel, Word
telnet

Bosch

2001-09-11T00:00:00+00:00

Embedded software technician on engine control units. I wrote and tested the engine functions that lived inside the Motronic ECU — code that had to be right long before it ever reached a car.

What I did🔗

Analyse and develop engine functions in C/C++ for Motronic
Review code for performance
Write unit and integration tests
Freeze versions before series production
Validate on test benches and on vehicles
Document each function
Write the test methodology

Tech stack🔗

C, C++, ASM
CodeWright
Messplatz Labcar (bench lab)
in-house tools and software

Towercast

2001-03-11T00:00:00+00:00

My first internship: building a small Windows application to read GPS data. It taught me a lot about serial links, data frames, and keeping an interface readable.

What I did🔗

Understand how the GPS system works
Receive NMEA data from a GARMIN GPS over an RS232 serial link
Parse and transform the NMEA frames
Program the reception in Delphi
Build a GUI to make the receiver output readable

Tech stack🔗

Delphi, Visual Studio
RS232 serial link
GPS NMEA data
GPS technology

/void !

DeepSeek Flash 4.1: the open source model putting pressure on the American players

Open source, for real🔗

And the price…🔗

Real competition for the American models🔗

What it changes for us🔗

Fair warning, though🔗

In short🔗

From « Code Monkey » to Systemic Orchestrator: how the developer job is changing

Less syntax, more judgment🔗

The big picture rather than the isolated brick🔗

What about beginners?🔗

In short🔗

Alacritty: my everyday terminal ⚡

Herdr: the terminal multiplexer built for AI agents 🤖

What is Herdr?🔗

What Herdr lets you do🔗

In short🔗

Criteo — SRE

What I did🔗

Tech stack🔗

Rust app in terminal 🖥️

fd🔗

exa🔗

dust🔗

alacritty🔗

spotify-tui🔗

🚧 Gitlab and Github pages handling the same code 🤓

Gitlab - Master🔗

create our static website repo🔗

create gh-pages branch🔗

add web stuff inside the repo🔗

configure CI🔗

Github - Mirror🔗

create our static website repo🔗

Branch configuration🔗

Git mirror🔗

configure CI🔗

Conclusion🔗

TheMove 🛴

Git two accounts alias

Bonus🔗

Criteo — Escalation Engineer

What I did🔗

Interview🔗

Tech stack🔗

Veolia — Treasury suite

What I did🔗

Analysis and delivery🔗

Continuous improvement🔗

Tech stack🔗

Caisse des dépôts

What I did🔗

Solutions analyst🔗

Business analyst🔗

Tech stack🔗

Veolia — Next treasury project

What I did🔗

Integration and environments🔗

TMA / L2-L3 support🔗

Tech stack🔗

Accor

What I did🔗

Tech stack🔗

Opéra National de Paris

What I did🔗

Tech stack🔗

Bosch

What I did🔗

Tech stack🔗

Towercast

What I did🔗

Tech stack🔗

`fd`🔗

`exa`🔗

`dust`🔗

`alacritty`🔗

`spotify-tui`🔗